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>
34 #include <linux/dma-buf.h>
36 #include <drm/drm_drv.h>
37 #include <drm/amdgpu_drm.h>
38 #include <drm/drm_cache.h>
40 #include "amdgpu_trace.h"
41 #include "amdgpu_amdkfd.h"
42 #include "amdgpu_vram_mgr.h"
47 * This defines the interfaces to operate on an &amdgpu_bo buffer object which
48 * represents memory used by driver (VRAM, system memory, etc.). The driver
49 * provides DRM/GEM APIs to userspace. DRM/GEM APIs then use these interfaces
50 * to create/destroy/set buffer object which are then managed by the kernel TTM
52 * The interfaces are also used internally by kernel clients, including gfx,
53 * uvd, etc. for kernel managed allocations used by the GPU.
57 static void amdgpu_bo_destroy(struct ttm_buffer_object *tbo)
59 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo);
63 if (bo->tbo.base.import_attach)
64 drm_prime_gem_destroy(&bo->tbo.base, bo->tbo.sg);
65 drm_gem_object_release(&bo->tbo.base);
66 amdgpu_bo_unref(&bo->parent);
70 static void amdgpu_bo_user_destroy(struct ttm_buffer_object *tbo)
72 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo);
73 struct amdgpu_bo_user *ubo;
75 ubo = to_amdgpu_bo_user(bo);
77 amdgpu_bo_destroy(tbo);
81 * amdgpu_bo_is_amdgpu_bo - check if the buffer object is an &amdgpu_bo
82 * @bo: buffer object to be checked
84 * Uses destroy function associated with the object to determine if this is
88 * true if the object belongs to &amdgpu_bo, false if not.
90 bool amdgpu_bo_is_amdgpu_bo(struct ttm_buffer_object *bo)
92 if (bo->destroy == &amdgpu_bo_destroy ||
93 bo->destroy == &amdgpu_bo_user_destroy)
100 * amdgpu_bo_placement_from_domain - set buffer's placement
101 * @abo: &amdgpu_bo buffer object whose placement is to be set
102 * @domain: requested domain
104 * Sets buffer's placement according to requested domain and the buffer's
107 void amdgpu_bo_placement_from_domain(struct amdgpu_bo *abo, u32 domain)
109 struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
110 struct ttm_placement *placement = &abo->placement;
111 struct ttm_place *places = abo->placements;
112 u64 flags = abo->flags;
115 if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
116 unsigned int visible_pfn = adev->gmc.visible_vram_size >> PAGE_SHIFT;
117 int8_t mem_id = KFD_XCP_MEM_ID(adev, abo->xcp_id);
119 if (adev->gmc.mem_partitions && mem_id >= 0) {
120 places[c].fpfn = adev->gmc.mem_partitions[mem_id].range.fpfn;
122 * memory partition range lpfn is inclusive start + size - 1
123 * TTM place lpfn is exclusive start + size
125 places[c].lpfn = adev->gmc.mem_partitions[mem_id].range.lpfn + 1;
130 places[c].mem_type = TTM_PL_VRAM;
133 if (flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
134 places[c].lpfn = min_not_zero(places[c].lpfn, visible_pfn);
136 places[c].flags |= TTM_PL_FLAG_TOPDOWN;
138 if (abo->tbo.type == ttm_bo_type_kernel &&
139 flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)
140 places[c].flags |= TTM_PL_FLAG_CONTIGUOUS;
145 if (domain & AMDGPU_GEM_DOMAIN_DOORBELL) {
148 places[c].mem_type = AMDGPU_PL_DOORBELL;
153 if (domain & AMDGPU_GEM_DOMAIN_GTT) {
157 abo->flags & AMDGPU_GEM_CREATE_PREEMPTIBLE ?
158 AMDGPU_PL_PREEMPT : TTM_PL_TT;
161 * When GTT is just an alternative to VRAM make sure that we
162 * only use it as fallback and still try to fill up VRAM first.
164 if (domain & abo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM &&
165 !(adev->flags & AMD_IS_APU))
166 places[c].flags |= TTM_PL_FLAG_FALLBACK;
170 if (domain & AMDGPU_GEM_DOMAIN_CPU) {
173 places[c].mem_type = TTM_PL_SYSTEM;
178 if (domain & AMDGPU_GEM_DOMAIN_GDS) {
181 places[c].mem_type = AMDGPU_PL_GDS;
186 if (domain & AMDGPU_GEM_DOMAIN_GWS) {
189 places[c].mem_type = AMDGPU_PL_GWS;
194 if (domain & AMDGPU_GEM_DOMAIN_OA) {
197 places[c].mem_type = AMDGPU_PL_OA;
205 places[c].mem_type = TTM_PL_SYSTEM;
210 BUG_ON(c > AMDGPU_BO_MAX_PLACEMENTS);
212 placement->num_placement = c;
213 placement->placement = places;
217 * amdgpu_bo_create_reserved - create reserved BO for kernel use
219 * @adev: amdgpu device object
220 * @size: size for the new BO
221 * @align: alignment for the new BO
222 * @domain: where to place it
223 * @bo_ptr: used to initialize BOs in structures
224 * @gpu_addr: GPU addr of the pinned BO
225 * @cpu_addr: optional CPU address mapping
227 * Allocates and pins a BO for kernel internal use, and returns it still
230 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
233 * 0 on success, negative error code otherwise.
235 int amdgpu_bo_create_reserved(struct amdgpu_device *adev,
236 unsigned long size, int align,
237 u32 domain, struct amdgpu_bo **bo_ptr,
238 u64 *gpu_addr, void **cpu_addr)
240 struct amdgpu_bo_param bp;
245 amdgpu_bo_unref(bo_ptr);
249 memset(&bp, 0, sizeof(bp));
251 bp.byte_align = align;
253 bp.flags = cpu_addr ? AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED
254 : AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
255 bp.flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
256 bp.type = ttm_bo_type_kernel;
258 bp.bo_ptr_size = sizeof(struct amdgpu_bo);
261 r = amdgpu_bo_create(adev, &bp, bo_ptr);
263 dev_err(adev->dev, "(%d) failed to allocate kernel bo\n",
270 r = amdgpu_bo_reserve(*bo_ptr, false);
272 dev_err(adev->dev, "(%d) failed to reserve kernel bo\n", r);
276 r = amdgpu_bo_pin(*bo_ptr, domain);
278 dev_err(adev->dev, "(%d) kernel bo pin failed\n", r);
279 goto error_unreserve;
282 r = amdgpu_ttm_alloc_gart(&(*bo_ptr)->tbo);
284 dev_err(adev->dev, "%p bind failed\n", *bo_ptr);
289 *gpu_addr = amdgpu_bo_gpu_offset(*bo_ptr);
292 r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
294 dev_err(adev->dev, "(%d) kernel bo map failed\n", r);
302 amdgpu_bo_unpin(*bo_ptr);
304 amdgpu_bo_unreserve(*bo_ptr);
308 amdgpu_bo_unref(bo_ptr);
314 * amdgpu_bo_create_kernel - create BO for kernel use
316 * @adev: amdgpu device object
317 * @size: size for the new BO
318 * @align: alignment for the new BO
319 * @domain: where to place it
320 * @bo_ptr: used to initialize BOs in structures
321 * @gpu_addr: GPU addr of the pinned BO
322 * @cpu_addr: optional CPU address mapping
324 * Allocates and pins a BO for kernel internal use.
326 * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
329 * 0 on success, negative error code otherwise.
331 int amdgpu_bo_create_kernel(struct amdgpu_device *adev,
332 unsigned long size, int align,
333 u32 domain, struct amdgpu_bo **bo_ptr,
334 u64 *gpu_addr, void **cpu_addr)
338 r = amdgpu_bo_create_reserved(adev, size, align, domain, bo_ptr,
345 amdgpu_bo_unreserve(*bo_ptr);
351 * amdgpu_bo_create_kernel_at - create BO for kernel use at specific location
353 * @adev: amdgpu device object
354 * @offset: offset of the BO
355 * @size: size of the BO
356 * @bo_ptr: used to initialize BOs in structures
357 * @cpu_addr: optional CPU address mapping
359 * Creates a kernel BO at a specific offset in VRAM.
362 * 0 on success, negative error code otherwise.
364 int amdgpu_bo_create_kernel_at(struct amdgpu_device *adev,
365 uint64_t offset, uint64_t size,
366 struct amdgpu_bo **bo_ptr, void **cpu_addr)
368 struct ttm_operation_ctx ctx = { false, false };
373 size = ALIGN(size, PAGE_SIZE);
375 r = amdgpu_bo_create_reserved(adev, size, PAGE_SIZE,
376 AMDGPU_GEM_DOMAIN_VRAM, bo_ptr, NULL,
381 if ((*bo_ptr) == NULL)
385 * Remove the original mem node and create a new one at the request
389 amdgpu_bo_kunmap(*bo_ptr);
391 ttm_resource_free(&(*bo_ptr)->tbo, &(*bo_ptr)->tbo.resource);
393 for (i = 0; i < (*bo_ptr)->placement.num_placement; ++i) {
394 (*bo_ptr)->placements[i].fpfn = offset >> PAGE_SHIFT;
395 (*bo_ptr)->placements[i].lpfn = (offset + size) >> PAGE_SHIFT;
397 r = ttm_bo_mem_space(&(*bo_ptr)->tbo, &(*bo_ptr)->placement,
398 &(*bo_ptr)->tbo.resource, &ctx);
403 r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
408 amdgpu_bo_unreserve(*bo_ptr);
412 amdgpu_bo_unreserve(*bo_ptr);
413 amdgpu_bo_unref(bo_ptr);
418 * amdgpu_bo_free_kernel - free BO for kernel use
420 * @bo: amdgpu BO to free
421 * @gpu_addr: pointer to where the BO's GPU memory space address was stored
422 * @cpu_addr: pointer to where the BO's CPU memory space address was stored
424 * unmaps and unpin a BO for kernel internal use.
426 void amdgpu_bo_free_kernel(struct amdgpu_bo **bo, u64 *gpu_addr,
432 WARN_ON(amdgpu_ttm_adev((*bo)->tbo.bdev)->in_suspend);
434 if (likely(amdgpu_bo_reserve(*bo, true) == 0)) {
436 amdgpu_bo_kunmap(*bo);
438 amdgpu_bo_unpin(*bo);
439 amdgpu_bo_unreserve(*bo);
450 /* Validate bo size is bit bigger than the request domain */
451 static bool amdgpu_bo_validate_size(struct amdgpu_device *adev,
452 unsigned long size, u32 domain)
454 struct ttm_resource_manager *man = NULL;
457 * If GTT is part of requested domains the check must succeed to
458 * allow fall back to GTT.
460 if (domain & AMDGPU_GEM_DOMAIN_GTT)
461 man = ttm_manager_type(&adev->mman.bdev, TTM_PL_TT);
462 else if (domain & AMDGPU_GEM_DOMAIN_VRAM)
463 man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM);
468 if (domain & AMDGPU_GEM_DOMAIN_GTT)
469 WARN_ON_ONCE("GTT domain requested but GTT mem manager uninitialized");
473 /* TODO add more domains checks, such as AMDGPU_GEM_DOMAIN_CPU, _DOMAIN_DOORBELL */
474 if (size < man->size)
477 DRM_DEBUG("BO size %lu > total memory in domain: %llu\n", size, man->size);
481 bool amdgpu_bo_support_uswc(u64 bo_flags)
485 /* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
486 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
489 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
490 /* Don't try to enable write-combining when it can't work, or things
492 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
495 #ifndef CONFIG_COMPILE_TEST
496 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
497 thanks to write-combining
500 if (bo_flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
501 DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
502 "better performance thanks to write-combining\n");
505 /* For architectures that don't support WC memory,
506 * mask out the WC flag from the BO
508 if (!drm_arch_can_wc_memory())
516 * amdgpu_bo_create - create an &amdgpu_bo buffer object
517 * @adev: amdgpu device object
518 * @bp: parameters to be used for the buffer object
519 * @bo_ptr: pointer to the buffer object pointer
521 * Creates an &amdgpu_bo buffer object.
524 * 0 for success or a negative error code on failure.
526 int amdgpu_bo_create(struct amdgpu_device *adev,
527 struct amdgpu_bo_param *bp,
528 struct amdgpu_bo **bo_ptr)
530 struct ttm_operation_ctx ctx = {
531 .interruptible = (bp->type != ttm_bo_type_kernel),
532 .no_wait_gpu = bp->no_wait_gpu,
533 /* We opt to avoid OOM on system pages allocations */
534 .gfp_retry_mayfail = true,
535 .allow_res_evict = bp->type != ttm_bo_type_kernel,
538 struct amdgpu_bo *bo;
539 unsigned long page_align, size = bp->size;
542 /* Note that GDS/GWS/OA allocates 1 page per byte/resource. */
543 if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA)) {
544 /* GWS and OA don't need any alignment. */
545 page_align = bp->byte_align;
548 } else if (bp->domain & AMDGPU_GEM_DOMAIN_GDS) {
549 /* Both size and alignment must be a multiple of 4. */
550 page_align = ALIGN(bp->byte_align, 4);
551 size = ALIGN(size, 4) << PAGE_SHIFT;
553 /* Memory should be aligned at least to a page size. */
554 page_align = ALIGN(bp->byte_align, PAGE_SIZE) >> PAGE_SHIFT;
555 size = ALIGN(size, PAGE_SIZE);
558 if (!amdgpu_bo_validate_size(adev, size, bp->domain))
561 BUG_ON(bp->bo_ptr_size < sizeof(struct amdgpu_bo));
564 bo = kvzalloc(bp->bo_ptr_size, GFP_KERNEL);
567 drm_gem_private_object_init(adev_to_drm(adev), &bo->tbo.base, size);
568 bo->tbo.base.funcs = &amdgpu_gem_object_funcs;
570 bo->preferred_domains = bp->preferred_domain ? bp->preferred_domain :
572 bo->allowed_domains = bo->preferred_domains;
573 if (bp->type != ttm_bo_type_kernel &&
574 !(bp->flags & AMDGPU_GEM_CREATE_DISCARDABLE) &&
575 bo->allowed_domains == AMDGPU_GEM_DOMAIN_VRAM)
576 bo->allowed_domains |= AMDGPU_GEM_DOMAIN_GTT;
578 bo->flags = bp->flags;
580 if (adev->gmc.mem_partitions)
581 /* For GPUs with spatial partitioning, bo->xcp_id=-1 means any partition */
582 bo->xcp_id = bp->xcp_id_plus1 - 1;
584 /* For GPUs without spatial partitioning */
587 if (!amdgpu_bo_support_uswc(bo->flags))
588 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
590 bo->tbo.bdev = &adev->mman.bdev;
591 if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA |
592 AMDGPU_GEM_DOMAIN_GDS))
593 amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
595 amdgpu_bo_placement_from_domain(bo, bp->domain);
596 if (bp->type == ttm_bo_type_kernel)
597 bo->tbo.priority = 2;
598 else if (!(bp->flags & AMDGPU_GEM_CREATE_DISCARDABLE))
599 bo->tbo.priority = 1;
602 bp->destroy = &amdgpu_bo_destroy;
604 r = ttm_bo_init_reserved(&adev->mman.bdev, &bo->tbo, bp->type,
605 &bo->placement, page_align, &ctx, NULL,
606 bp->resv, bp->destroy);
607 if (unlikely(r != 0))
610 if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
611 amdgpu_res_cpu_visible(adev, bo->tbo.resource))
612 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved,
615 amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved, 0);
617 if (bp->flags & AMDGPU_GEM_CREATE_VRAM_CLEARED &&
618 bo->tbo.resource->mem_type == TTM_PL_VRAM) {
619 struct dma_fence *fence;
621 r = amdgpu_ttm_clear_buffer(bo, bo->tbo.base.resv, &fence);
625 dma_resv_add_fence(bo->tbo.base.resv, fence,
626 DMA_RESV_USAGE_KERNEL);
627 dma_fence_put(fence);
630 amdgpu_bo_unreserve(bo);
633 trace_amdgpu_bo_create(bo);
635 /* Treat CPU_ACCESS_REQUIRED only as a hint if given by UMD */
636 if (bp->type == ttm_bo_type_device)
637 bo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
643 dma_resv_unlock(bo->tbo.base.resv);
644 amdgpu_bo_unref(&bo);
649 * amdgpu_bo_create_user - create an &amdgpu_bo_user buffer object
650 * @adev: amdgpu device object
651 * @bp: parameters to be used for the buffer object
652 * @ubo_ptr: pointer to the buffer object pointer
654 * Create a BO to be used by user application;
657 * 0 for success or a negative error code on failure.
660 int amdgpu_bo_create_user(struct amdgpu_device *adev,
661 struct amdgpu_bo_param *bp,
662 struct amdgpu_bo_user **ubo_ptr)
664 struct amdgpu_bo *bo_ptr;
667 bp->bo_ptr_size = sizeof(struct amdgpu_bo_user);
668 bp->destroy = &amdgpu_bo_user_destroy;
669 r = amdgpu_bo_create(adev, bp, &bo_ptr);
673 *ubo_ptr = to_amdgpu_bo_user(bo_ptr);
678 * amdgpu_bo_create_vm - create an &amdgpu_bo_vm buffer object
679 * @adev: amdgpu device object
680 * @bp: parameters to be used for the buffer object
681 * @vmbo_ptr: pointer to the buffer object pointer
683 * Create a BO to be for GPUVM.
686 * 0 for success or a negative error code on failure.
689 int amdgpu_bo_create_vm(struct amdgpu_device *adev,
690 struct amdgpu_bo_param *bp,
691 struct amdgpu_bo_vm **vmbo_ptr)
693 struct amdgpu_bo *bo_ptr;
696 /* bo_ptr_size will be determined by the caller and it depends on
697 * num of amdgpu_vm_pt entries.
699 BUG_ON(bp->bo_ptr_size < sizeof(struct amdgpu_bo_vm));
700 r = amdgpu_bo_create(adev, bp, &bo_ptr);
704 *vmbo_ptr = to_amdgpu_bo_vm(bo_ptr);
709 * amdgpu_bo_kmap - map an &amdgpu_bo buffer object
710 * @bo: &amdgpu_bo buffer object to be mapped
711 * @ptr: kernel virtual address to be returned
713 * Calls ttm_bo_kmap() to set up the kernel virtual mapping; calls
714 * amdgpu_bo_kptr() to get the kernel virtual address.
717 * 0 for success or a negative error code on failure.
719 int amdgpu_bo_kmap(struct amdgpu_bo *bo, void **ptr)
724 if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
727 r = dma_resv_wait_timeout(bo->tbo.base.resv, DMA_RESV_USAGE_KERNEL,
728 false, MAX_SCHEDULE_TIMEOUT);
732 kptr = amdgpu_bo_kptr(bo);
739 r = ttm_bo_kmap(&bo->tbo, 0, PFN_UP(bo->tbo.base.size), &bo->kmap);
744 *ptr = amdgpu_bo_kptr(bo);
750 * amdgpu_bo_kptr - returns a kernel virtual address of the buffer object
751 * @bo: &amdgpu_bo buffer object
753 * Calls ttm_kmap_obj_virtual() to get the kernel virtual address
756 * the virtual address of a buffer object area.
758 void *amdgpu_bo_kptr(struct amdgpu_bo *bo)
762 return ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
766 * amdgpu_bo_kunmap - unmap an &amdgpu_bo buffer object
767 * @bo: &amdgpu_bo buffer object to be unmapped
769 * Unmaps a kernel map set up by amdgpu_bo_kmap().
771 void amdgpu_bo_kunmap(struct amdgpu_bo *bo)
774 ttm_bo_kunmap(&bo->kmap);
778 * amdgpu_bo_ref - reference an &amdgpu_bo buffer object
779 * @bo: &amdgpu_bo buffer object
781 * References the contained &ttm_buffer_object.
784 * a refcounted pointer to the &amdgpu_bo buffer object.
786 struct amdgpu_bo *amdgpu_bo_ref(struct amdgpu_bo *bo)
791 drm_gem_object_get(&bo->tbo.base);
796 * amdgpu_bo_unref - unreference an &amdgpu_bo buffer object
797 * @bo: &amdgpu_bo buffer object
799 * Unreferences the contained &ttm_buffer_object and clear the pointer
801 void amdgpu_bo_unref(struct amdgpu_bo **bo)
806 drm_gem_object_put(&(*bo)->tbo.base);
811 * amdgpu_bo_pin - pin an &amdgpu_bo buffer object
812 * @bo: &amdgpu_bo buffer object to be pinned
813 * @domain: domain to be pinned to
815 * Pins the buffer object according to requested domain. If the memory is
816 * unbound gart memory, binds the pages into gart table. Adjusts pin_count and
817 * pin_size accordingly.
819 * Pinning means to lock pages in memory along with keeping them at a fixed
820 * offset. It is required when a buffer can not be moved, for example, when
821 * a display buffer is being scanned out.
824 * 0 for success or a negative error code on failure.
826 int amdgpu_bo_pin(struct amdgpu_bo *bo, u32 domain)
828 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
829 struct ttm_operation_ctx ctx = { false, false };
832 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm))
835 /* Check domain to be pinned to against preferred domains */
836 if (bo->preferred_domains & domain)
837 domain = bo->preferred_domains & domain;
839 /* A shared bo cannot be migrated to VRAM */
840 if (bo->tbo.base.import_attach) {
841 if (domain & AMDGPU_GEM_DOMAIN_GTT)
842 domain = AMDGPU_GEM_DOMAIN_GTT;
847 if (bo->tbo.pin_count) {
848 uint32_t mem_type = bo->tbo.resource->mem_type;
849 uint32_t mem_flags = bo->tbo.resource->placement;
851 if (!(domain & amdgpu_mem_type_to_domain(mem_type)))
854 if ((mem_type == TTM_PL_VRAM) &&
855 (bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS) &&
856 !(mem_flags & TTM_PL_FLAG_CONTIGUOUS))
859 ttm_bo_pin(&bo->tbo);
863 /* This assumes only APU display buffers are pinned with (VRAM|GTT).
864 * See function amdgpu_display_supported_domains()
866 domain = amdgpu_bo_get_preferred_domain(adev, domain);
868 if (bo->tbo.base.import_attach)
869 dma_buf_pin(bo->tbo.base.import_attach);
871 /* force to pin into visible video ram */
872 if (!(bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS))
873 bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
874 amdgpu_bo_placement_from_domain(bo, domain);
875 for (i = 0; i < bo->placement.num_placement; i++) {
876 if (bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS &&
877 bo->placements[i].mem_type == TTM_PL_VRAM)
878 bo->placements[i].flags |= TTM_PL_FLAG_CONTIGUOUS;
881 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
883 dev_err(adev->dev, "%p pin failed\n", bo);
887 ttm_bo_pin(&bo->tbo);
889 if (bo->tbo.resource->mem_type == TTM_PL_VRAM) {
890 atomic64_add(amdgpu_bo_size(bo), &adev->vram_pin_size);
891 atomic64_add(amdgpu_vram_mgr_bo_visible_size(bo),
892 &adev->visible_pin_size);
893 } else if (bo->tbo.resource->mem_type == TTM_PL_TT) {
894 atomic64_add(amdgpu_bo_size(bo), &adev->gart_pin_size);
902 * amdgpu_bo_unpin - unpin an &amdgpu_bo buffer object
903 * @bo: &amdgpu_bo buffer object to be unpinned
905 * Decreases the pin_count, and clears the flags if pin_count reaches 0.
906 * Changes placement and pin size accordingly.
909 * 0 for success or a negative error code on failure.
911 void amdgpu_bo_unpin(struct amdgpu_bo *bo)
913 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
915 ttm_bo_unpin(&bo->tbo);
916 if (bo->tbo.pin_count)
919 if (bo->tbo.base.import_attach)
920 dma_buf_unpin(bo->tbo.base.import_attach);
922 if (bo->tbo.resource->mem_type == TTM_PL_VRAM) {
923 atomic64_sub(amdgpu_bo_size(bo), &adev->vram_pin_size);
924 atomic64_sub(amdgpu_vram_mgr_bo_visible_size(bo),
925 &adev->visible_pin_size);
926 } else if (bo->tbo.resource->mem_type == TTM_PL_TT) {
927 atomic64_sub(amdgpu_bo_size(bo), &adev->gart_pin_size);
932 static const char * const amdgpu_vram_names[] = {
949 * amdgpu_bo_init - initialize memory manager
950 * @adev: amdgpu device object
952 * Calls amdgpu_ttm_init() to initialize amdgpu memory manager.
955 * 0 for success or a negative error code on failure.
957 int amdgpu_bo_init(struct amdgpu_device *adev)
959 /* On A+A platform, VRAM can be mapped as WB */
960 if (!adev->gmc.xgmi.connected_to_cpu && !adev->gmc.is_app_apu) {
961 /* reserve PAT memory space to WC for VRAM */
962 int r = arch_io_reserve_memtype_wc(adev->gmc.aper_base,
963 adev->gmc.aper_size);
966 DRM_ERROR("Unable to set WC memtype for the aperture base\n");
970 /* Add an MTRR for the VRAM */
971 adev->gmc.vram_mtrr = arch_phys_wc_add(adev->gmc.aper_base,
972 adev->gmc.aper_size);
975 DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
976 adev->gmc.mc_vram_size >> 20,
977 (unsigned long long)adev->gmc.aper_size >> 20);
978 DRM_INFO("RAM width %dbits %s\n",
979 adev->gmc.vram_width, amdgpu_vram_names[adev->gmc.vram_type]);
980 return amdgpu_ttm_init(adev);
984 * amdgpu_bo_fini - tear down memory manager
985 * @adev: amdgpu device object
987 * Reverses amdgpu_bo_init() to tear down memory manager.
989 void amdgpu_bo_fini(struct amdgpu_device *adev)
993 amdgpu_ttm_fini(adev);
995 if (drm_dev_enter(adev_to_drm(adev), &idx)) {
996 if (!adev->gmc.xgmi.connected_to_cpu && !adev->gmc.is_app_apu) {
997 arch_phys_wc_del(adev->gmc.vram_mtrr);
998 arch_io_free_memtype_wc(adev->gmc.aper_base, adev->gmc.aper_size);
1005 * amdgpu_bo_set_tiling_flags - set tiling flags
1006 * @bo: &amdgpu_bo buffer object
1007 * @tiling_flags: new flags
1009 * Sets buffer object's tiling flags with the new one. Used by GEM ioctl or
1010 * kernel driver to set the tiling flags on a buffer.
1013 * 0 for success or a negative error code on failure.
1015 int amdgpu_bo_set_tiling_flags(struct amdgpu_bo *bo, u64 tiling_flags)
1017 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1018 struct amdgpu_bo_user *ubo;
1020 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1021 if (adev->family <= AMDGPU_FAMILY_CZ &&
1022 AMDGPU_TILING_GET(tiling_flags, TILE_SPLIT) > 6)
1025 ubo = to_amdgpu_bo_user(bo);
1026 ubo->tiling_flags = tiling_flags;
1031 * amdgpu_bo_get_tiling_flags - get tiling flags
1032 * @bo: &amdgpu_bo buffer object
1033 * @tiling_flags: returned flags
1035 * Gets buffer object's tiling flags. Used by GEM ioctl or kernel driver to
1036 * set the tiling flags on a buffer.
1038 void amdgpu_bo_get_tiling_flags(struct amdgpu_bo *bo, u64 *tiling_flags)
1040 struct amdgpu_bo_user *ubo;
1042 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1043 dma_resv_assert_held(bo->tbo.base.resv);
1044 ubo = to_amdgpu_bo_user(bo);
1047 *tiling_flags = ubo->tiling_flags;
1051 * amdgpu_bo_set_metadata - set metadata
1052 * @bo: &amdgpu_bo buffer object
1053 * @metadata: new metadata
1054 * @metadata_size: size of the new metadata
1055 * @flags: flags of the new metadata
1057 * Sets buffer object's metadata, its size and flags.
1058 * Used via GEM ioctl.
1061 * 0 for success or a negative error code on failure.
1063 int amdgpu_bo_set_metadata(struct amdgpu_bo *bo, void *metadata,
1064 u32 metadata_size, uint64_t flags)
1066 struct amdgpu_bo_user *ubo;
1069 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1070 ubo = to_amdgpu_bo_user(bo);
1071 if (!metadata_size) {
1072 if (ubo->metadata_size) {
1073 kfree(ubo->metadata);
1074 ubo->metadata = NULL;
1075 ubo->metadata_size = 0;
1080 if (metadata == NULL)
1083 buffer = kmemdup(metadata, metadata_size, GFP_KERNEL);
1087 kfree(ubo->metadata);
1088 ubo->metadata_flags = flags;
1089 ubo->metadata = buffer;
1090 ubo->metadata_size = metadata_size;
1096 * amdgpu_bo_get_metadata - get metadata
1097 * @bo: &amdgpu_bo buffer object
1098 * @buffer: returned metadata
1099 * @buffer_size: size of the buffer
1100 * @metadata_size: size of the returned metadata
1101 * @flags: flags of the returned metadata
1103 * Gets buffer object's metadata, its size and flags. buffer_size shall not be
1104 * less than metadata_size.
1105 * Used via GEM ioctl.
1108 * 0 for success or a negative error code on failure.
1110 int amdgpu_bo_get_metadata(struct amdgpu_bo *bo, void *buffer,
1111 size_t buffer_size, uint32_t *metadata_size,
1114 struct amdgpu_bo_user *ubo;
1116 if (!buffer && !metadata_size)
1119 BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1120 ubo = to_amdgpu_bo_user(bo);
1122 *metadata_size = ubo->metadata_size;
1125 if (buffer_size < ubo->metadata_size)
1128 if (ubo->metadata_size)
1129 memcpy(buffer, ubo->metadata, ubo->metadata_size);
1133 *flags = ubo->metadata_flags;
1139 * amdgpu_bo_move_notify - notification about a memory move
1140 * @bo: pointer to a buffer object
1141 * @evict: if this move is evicting the buffer from the graphics address space
1142 * @new_mem: new resource for backing the BO
1144 * Marks the corresponding &amdgpu_bo buffer object as invalid, also performs
1146 * TTM driver callback which is called when ttm moves a buffer.
1148 void amdgpu_bo_move_notify(struct ttm_buffer_object *bo,
1150 struct ttm_resource *new_mem)
1152 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1153 struct ttm_resource *old_mem = bo->resource;
1154 struct amdgpu_bo *abo;
1156 if (!amdgpu_bo_is_amdgpu_bo(bo))
1159 abo = ttm_to_amdgpu_bo(bo);
1160 amdgpu_vm_bo_invalidate(adev, abo, evict);
1162 amdgpu_bo_kunmap(abo);
1164 if (abo->tbo.base.dma_buf && !abo->tbo.base.import_attach &&
1165 old_mem && old_mem->mem_type != TTM_PL_SYSTEM)
1166 dma_buf_move_notify(abo->tbo.base.dma_buf);
1168 /* move_notify is called before move happens */
1169 trace_amdgpu_bo_move(abo, new_mem ? new_mem->mem_type : -1,
1170 old_mem ? old_mem->mem_type : -1);
1173 void amdgpu_bo_get_memory(struct amdgpu_bo *bo,
1174 struct amdgpu_mem_stats *stats)
1176 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1177 struct ttm_resource *res = bo->tbo.resource;
1178 uint64_t size = amdgpu_bo_size(bo);
1179 struct drm_gem_object *obj;
1182 /* Abort if the BO doesn't currently have a backing store */
1186 obj = &bo->tbo.base;
1187 shared = drm_gem_object_is_shared_for_memory_stats(obj);
1189 switch (res->mem_type) {
1191 stats->vram += size;
1192 if (amdgpu_res_cpu_visible(adev, res))
1193 stats->visible_vram += size;
1195 stats->vram_shared += size;
1200 stats->gtt_shared += size;
1206 stats->cpu_shared += size;
1210 if (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) {
1211 stats->requested_vram += size;
1212 if (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
1213 stats->requested_visible_vram += size;
1215 if (res->mem_type != TTM_PL_VRAM) {
1216 stats->evicted_vram += size;
1217 if (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
1218 stats->evicted_visible_vram += size;
1220 } else if (bo->preferred_domains & AMDGPU_GEM_DOMAIN_GTT) {
1221 stats->requested_gtt += size;
1226 * amdgpu_bo_release_notify - notification about a BO being released
1227 * @bo: pointer to a buffer object
1229 * Wipes VRAM buffers whose contents should not be leaked before the
1230 * memory is released.
1232 void amdgpu_bo_release_notify(struct ttm_buffer_object *bo)
1234 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1235 struct dma_fence *fence = NULL;
1236 struct amdgpu_bo *abo;
1239 if (!amdgpu_bo_is_amdgpu_bo(bo))
1242 abo = ttm_to_amdgpu_bo(bo);
1244 WARN_ON(abo->vm_bo);
1247 amdgpu_amdkfd_release_notify(abo);
1249 /* We only remove the fence if the resv has individualized. */
1250 WARN_ON_ONCE(bo->type == ttm_bo_type_kernel
1251 && bo->base.resv != &bo->base._resv);
1252 if (bo->base.resv == &bo->base._resv)
1253 amdgpu_amdkfd_remove_fence_on_pt_pd_bos(abo);
1255 if (!bo->resource || bo->resource->mem_type != TTM_PL_VRAM ||
1256 !(abo->flags & AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE) ||
1257 adev->in_suspend || drm_dev_is_unplugged(adev_to_drm(adev)))
1260 if (WARN_ON_ONCE(!dma_resv_trylock(bo->base.resv)))
1263 r = amdgpu_fill_buffer(abo, 0, bo->base.resv, &fence, true);
1265 amdgpu_vram_mgr_set_cleared(bo->resource);
1266 amdgpu_bo_fence(abo, fence, false);
1267 dma_fence_put(fence);
1270 dma_resv_unlock(bo->base.resv);
1274 * amdgpu_bo_fault_reserve_notify - notification about a memory fault
1275 * @bo: pointer to a buffer object
1277 * Notifies the driver we are taking a fault on this BO and have reserved it,
1278 * also performs bookkeeping.
1279 * TTM driver callback for dealing with vm faults.
1282 * 0 for success or a negative error code on failure.
1284 vm_fault_t amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
1286 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1287 struct ttm_operation_ctx ctx = { false, false };
1288 struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
1291 /* Remember that this BO was accessed by the CPU */
1292 abo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
1294 if (amdgpu_res_cpu_visible(adev, bo->resource))
1297 /* Can't move a pinned BO to visible VRAM */
1298 if (abo->tbo.pin_count > 0)
1299 return VM_FAULT_SIGBUS;
1301 /* hurrah the memory is not visible ! */
1302 atomic64_inc(&adev->num_vram_cpu_page_faults);
1303 amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM |
1304 AMDGPU_GEM_DOMAIN_GTT);
1306 /* Avoid costly evictions; only set GTT as a busy placement */
1307 abo->placements[0].flags |= TTM_PL_FLAG_DESIRED;
1309 r = ttm_bo_validate(bo, &abo->placement, &ctx);
1310 if (unlikely(r == -EBUSY || r == -ERESTARTSYS))
1311 return VM_FAULT_NOPAGE;
1312 else if (unlikely(r))
1313 return VM_FAULT_SIGBUS;
1315 /* this should never happen */
1316 if (bo->resource->mem_type == TTM_PL_VRAM &&
1317 !amdgpu_res_cpu_visible(adev, bo->resource))
1318 return VM_FAULT_SIGBUS;
1320 ttm_bo_move_to_lru_tail_unlocked(bo);
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 dma_resv *resv = bo->tbo.base.resv;
1338 r = dma_resv_reserve_fences(resv, 1);
1340 /* As last resort on OOM we block for the fence */
1341 dma_fence_wait(fence, false);
1345 dma_resv_add_fence(resv, fence, shared ? DMA_RESV_USAGE_READ :
1346 DMA_RESV_USAGE_WRITE);
1350 * amdgpu_bo_sync_wait_resv - Wait for BO reservation fences
1352 * @adev: amdgpu device pointer
1353 * @resv: reservation object to sync to
1354 * @sync_mode: synchronization mode
1355 * @owner: fence owner
1356 * @intr: Whether the wait is interruptible
1358 * Extract the fences from the reservation object and waits for them to finish.
1361 * 0 on success, errno otherwise.
1363 int amdgpu_bo_sync_wait_resv(struct amdgpu_device *adev, struct dma_resv *resv,
1364 enum amdgpu_sync_mode sync_mode, void *owner,
1367 struct amdgpu_sync sync;
1370 amdgpu_sync_create(&sync);
1371 amdgpu_sync_resv(adev, &sync, resv, sync_mode, owner);
1372 r = amdgpu_sync_wait(&sync, intr);
1373 amdgpu_sync_free(&sync);
1378 * amdgpu_bo_sync_wait - Wrapper for amdgpu_bo_sync_wait_resv
1379 * @bo: buffer object to wait for
1380 * @owner: fence owner
1381 * @intr: Whether the wait is interruptible
1383 * Wrapper to wait for fences in a BO.
1385 * 0 on success, errno otherwise.
1387 int amdgpu_bo_sync_wait(struct amdgpu_bo *bo, void *owner, bool intr)
1389 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1391 return amdgpu_bo_sync_wait_resv(adev, bo->tbo.base.resv,
1392 AMDGPU_SYNC_NE_OWNER, owner, intr);
1396 * amdgpu_bo_gpu_offset - return GPU offset of bo
1397 * @bo: amdgpu object for which we query the offset
1399 * Note: object should either be pinned or reserved when calling this
1400 * function, it might be useful to add check for this for debugging.
1403 * current GPU offset of the object.
1405 u64 amdgpu_bo_gpu_offset(struct amdgpu_bo *bo)
1407 WARN_ON_ONCE(bo->tbo.resource->mem_type == TTM_PL_SYSTEM);
1408 WARN_ON_ONCE(!dma_resv_is_locked(bo->tbo.base.resv) &&
1409 !bo->tbo.pin_count && bo->tbo.type != ttm_bo_type_kernel);
1410 WARN_ON_ONCE(bo->tbo.resource->start == AMDGPU_BO_INVALID_OFFSET);
1411 WARN_ON_ONCE(bo->tbo.resource->mem_type == TTM_PL_VRAM &&
1412 !(bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS));
1414 return amdgpu_bo_gpu_offset_no_check(bo);
1418 * amdgpu_bo_gpu_offset_no_check - return GPU offset of bo
1419 * @bo: amdgpu object for which we query the offset
1422 * current GPU offset of the object without raising warnings.
1424 u64 amdgpu_bo_gpu_offset_no_check(struct amdgpu_bo *bo)
1426 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1427 uint64_t offset = AMDGPU_BO_INVALID_OFFSET;
1429 if (bo->tbo.resource->mem_type == TTM_PL_TT)
1430 offset = amdgpu_gmc_agp_addr(&bo->tbo);
1432 if (offset == AMDGPU_BO_INVALID_OFFSET)
1433 offset = (bo->tbo.resource->start << PAGE_SHIFT) +
1434 amdgpu_ttm_domain_start(adev, bo->tbo.resource->mem_type);
1436 return amdgpu_gmc_sign_extend(offset);
1440 * amdgpu_bo_get_preferred_domain - get preferred domain
1441 * @adev: amdgpu device object
1442 * @domain: allowed :ref:`memory domains <amdgpu_memory_domains>`
1445 * Which of the allowed domains is preferred for allocating the BO.
1447 uint32_t amdgpu_bo_get_preferred_domain(struct amdgpu_device *adev,
1450 if ((domain == (AMDGPU_GEM_DOMAIN_VRAM | AMDGPU_GEM_DOMAIN_GTT)) &&
1451 ((adev->asic_type == CHIP_CARRIZO) || (adev->asic_type == CHIP_STONEY))) {
1452 domain = AMDGPU_GEM_DOMAIN_VRAM;
1453 if (adev->gmc.real_vram_size <= AMDGPU_SG_THRESHOLD)
1454 domain = AMDGPU_GEM_DOMAIN_GTT;
1459 #if defined(CONFIG_DEBUG_FS)
1460 #define amdgpu_bo_print_flag(m, bo, flag) \
1462 if (bo->flags & (AMDGPU_GEM_CREATE_ ## flag)) { \
1463 seq_printf((m), " " #flag); \
1468 * amdgpu_bo_print_info - print BO info in debugfs file
1470 * @id: Index or Id of the BO
1471 * @bo: Requested BO for printing info
1474 * Print BO information in debugfs file
1477 * Size of the BO in bytes.
1479 u64 amdgpu_bo_print_info(int id, struct amdgpu_bo *bo, struct seq_file *m)
1481 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1482 struct dma_buf_attachment *attachment;
1483 struct dma_buf *dma_buf;
1484 const char *placement;
1485 unsigned int pin_count;
1488 if (dma_resv_trylock(bo->tbo.base.resv)) {
1489 if (!bo->tbo.resource) {
1492 switch (bo->tbo.resource->mem_type) {
1494 if (amdgpu_res_cpu_visible(adev, bo->tbo.resource))
1495 placement = "VRAM VISIBLE";
1511 case AMDGPU_PL_PREEMPT:
1512 placement = "PREEMPTIBLE";
1514 case AMDGPU_PL_DOORBELL:
1515 placement = "DOORBELL";
1523 dma_resv_unlock(bo->tbo.base.resv);
1525 placement = "UNKNOWN";
1528 size = amdgpu_bo_size(bo);
1529 seq_printf(m, "\t\t0x%08x: %12lld byte %s",
1530 id, size, placement);
1532 pin_count = READ_ONCE(bo->tbo.pin_count);
1534 seq_printf(m, " pin count %d", pin_count);
1536 dma_buf = READ_ONCE(bo->tbo.base.dma_buf);
1537 attachment = READ_ONCE(bo->tbo.base.import_attach);
1540 seq_printf(m, " imported from ino:%lu", file_inode(dma_buf->file)->i_ino);
1542 seq_printf(m, " exported as ino:%lu", file_inode(dma_buf->file)->i_ino);
1544 amdgpu_bo_print_flag(m, bo, CPU_ACCESS_REQUIRED);
1545 amdgpu_bo_print_flag(m, bo, NO_CPU_ACCESS);
1546 amdgpu_bo_print_flag(m, bo, CPU_GTT_USWC);
1547 amdgpu_bo_print_flag(m, bo, VRAM_CLEARED);
1548 amdgpu_bo_print_flag(m, bo, VRAM_CONTIGUOUS);
1549 amdgpu_bo_print_flag(m, bo, VM_ALWAYS_VALID);
1550 amdgpu_bo_print_flag(m, bo, EXPLICIT_SYNC);