Lines Matching +full:gfx +full:- +full:mem
1 // SPDX-License-Identifier: MIT
3 * Copyright 2014-2018 Advanced Micro Devices, Inc.
23 #include <linux/dma-buf.h>
74 #define domain_string(domain) domain_bit_to_string[ffs(domain)-1]
79 struct kgd_mem *mem) in kfd_mem_is_attached() argument
83 list_for_each_entry(entry, &mem->attachments, list) in kfd_mem_is_attached()
84 if (entry->bo_va->base.vm == avm) in kfd_mem_is_attached()
91 * reuse_dmamap() - Check whether adev can share the original
105 return (adev->ram_is_direct_mapped && bo_adev->ram_is_direct_mapped) || in reuse_dmamap()
106 (adev->dev->iommu_group == bo_adev->dev->iommu_group); in reuse_dmamap()
110 * System (TTM + userptr) memory - 15/16th System RAM
111 * TTM memory - 3/8th System RAM
116 uint64_t mem; in amdgpu_amdkfd_gpuvm_init_mem_limits() local
122 mem = si.totalram - si.totalhigh; in amdgpu_amdkfd_gpuvm_init_mem_limits()
123 mem *= si.mem_unit; in amdgpu_amdkfd_gpuvm_init_mem_limits()
126 kfd_mem_limit.max_system_mem_limit = mem - (mem >> 6); in amdgpu_amdkfd_gpuvm_init_mem_limits()
130 kfd_mem_limit.max_system_mem_limit -= AMDGPU_RESERVE_MEM_LIMIT; in amdgpu_amdkfd_gpuvm_init_mem_limits()
157 * amdgpu_amdkfd_reserve_mem_limit() - Decrease available memory by size
168 * returns -ENOMEM in case of error, ZERO otherwise
193 * For GFX 9.4.3, get the VRAM size from XCP structs in amdgpu_amdkfd_reserve_mem_limit()
196 return -EINVAL; in amdgpu_amdkfd_reserve_mem_limit()
199 if (adev->gmc.is_app_apu) { in amdgpu_amdkfd_reserve_mem_limit()
209 return -ENOMEM; in amdgpu_amdkfd_reserve_mem_limit()
222 (adev && xcp_id >= 0 && adev->kfd.vram_used[xcp_id] + vram_needed > in amdgpu_amdkfd_reserve_mem_limit()
223 vram_size - reserved_for_pt)) { in amdgpu_amdkfd_reserve_mem_limit()
224 ret = -ENOMEM; in amdgpu_amdkfd_reserve_mem_limit()
234 adev->kfd.vram_used[xcp_id] += vram_needed; in amdgpu_amdkfd_reserve_mem_limit()
235 adev->kfd.vram_used_aligned[xcp_id] += adev->gmc.is_app_apu ? in amdgpu_amdkfd_reserve_mem_limit()
253 kfd_mem_limit.system_mem_used -= size; in amdgpu_amdkfd_unreserve_mem_limit()
254 kfd_mem_limit.ttm_mem_used -= size; in amdgpu_amdkfd_unreserve_mem_limit()
257 "adev reference can't be null when alloc mem flags vram is set"); in amdgpu_amdkfd_unreserve_mem_limit()
262 adev->kfd.vram_used[xcp_id] -= size; in amdgpu_amdkfd_unreserve_mem_limit()
263 if (adev->gmc.is_app_apu) { in amdgpu_amdkfd_unreserve_mem_limit()
264 adev->kfd.vram_used_aligned[xcp_id] -= size; in amdgpu_amdkfd_unreserve_mem_limit()
265 kfd_mem_limit.system_mem_used -= size; in amdgpu_amdkfd_unreserve_mem_limit()
266 kfd_mem_limit.ttm_mem_used -= size; in amdgpu_amdkfd_unreserve_mem_limit()
268 adev->kfd.vram_used_aligned[xcp_id] -= in amdgpu_amdkfd_unreserve_mem_limit()
273 kfd_mem_limit.system_mem_used -= size; in amdgpu_amdkfd_unreserve_mem_limit()
280 WARN_ONCE(adev && xcp_id >= 0 && adev->kfd.vram_used[xcp_id] < 0, in amdgpu_amdkfd_unreserve_mem_limit()
293 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev); in amdgpu_amdkfd_release_notify()
294 u32 alloc_flags = bo->kfd_bo->alloc_flags; in amdgpu_amdkfd_release_notify()
298 bo->xcp_id); in amdgpu_amdkfd_release_notify()
300 kfree(bo->kfd_bo); in amdgpu_amdkfd_release_notify()
304 * create_dmamap_sg_bo() - Creates a amdgpu_bo object to reflect information
308 * @mem: BO of peer device that is being DMA mapped. Provides parameters
314 struct kgd_mem *mem, struct amdgpu_bo **bo_out) in create_dmamap_sg_bo() argument
320 ret = amdgpu_bo_reserve(mem->bo, false); in create_dmamap_sg_bo()
324 if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) in create_dmamap_sg_bo()
325 flags |= mem->bo->flags & (AMDGPU_GEM_CREATE_COHERENT | in create_dmamap_sg_bo()
328 ret = amdgpu_gem_object_create(adev, mem->bo->tbo.base.size, 1, in create_dmamap_sg_bo()
330 ttm_bo_type_sg, mem->bo->tbo.base.resv, &gem_obj, 0); in create_dmamap_sg_bo()
332 amdgpu_bo_unreserve(mem->bo); in create_dmamap_sg_bo()
336 return -EINVAL; in create_dmamap_sg_bo()
340 (*bo_out)->parent = amdgpu_bo_ref(mem->bo); in create_dmamap_sg_bo()
344 /* amdgpu_amdkfd_remove_eviction_fence - Removes eviction fence from BO's
351 * NOTE: Must be called with BO reserved i.e. bo->tbo.resv->lock held.
359 return -EINVAL; in amdgpu_amdkfd_remove_eviction_fence()
365 dma_resv_replace_fences(bo->tbo.base.resv, ef->base.context, in amdgpu_amdkfd_remove_eviction_fence()
381 while (root->parent) in amdgpu_amdkfd_remove_fence_on_pt_pd_bos()
382 root = root->parent; in amdgpu_amdkfd_remove_fence_on_pt_pd_bos()
384 vm_bo = root->vm_bo; in amdgpu_amdkfd_remove_fence_on_pt_pd_bos()
388 vm = vm_bo->vm; in amdgpu_amdkfd_remove_fence_on_pt_pd_bos()
392 info = vm->process_info; in amdgpu_amdkfd_remove_fence_on_pt_pd_bos()
393 if (!info || !info->eviction_fence) in amdgpu_amdkfd_remove_fence_on_pt_pd_bos()
396 ef = container_of(dma_fence_get(&info->eviction_fence->base), in amdgpu_amdkfd_remove_fence_on_pt_pd_bos()
399 BUG_ON(!dma_resv_trylock(bo->tbo.base.resv)); in amdgpu_amdkfd_remove_fence_on_pt_pd_bos()
401 dma_resv_unlock(bo->tbo.base.resv); in amdgpu_amdkfd_remove_fence_on_pt_pd_bos()
403 dma_fence_put(&ef->base); in amdgpu_amdkfd_remove_fence_on_pt_pd_bos()
413 if (WARN(amdgpu_ttm_tt_get_usermm(bo->tbo.ttm), in amdgpu_amdkfd_bo_validate()
415 return -EINVAL; in amdgpu_amdkfd_bo_validate()
419 ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); in amdgpu_amdkfd_bo_validate()
442 ret = dma_resv_reserve_fences(bo->tbo.base.resv, 1); in amdgpu_amdkfd_bo_validate_and_fence()
446 dma_resv_add_fence(bo->tbo.base.resv, fence, in amdgpu_amdkfd_bo_validate_and_fence()
457 return amdgpu_amdkfd_bo_validate(bo, bo->allowed_domains, false); in amdgpu_amdkfd_validate_vm_bo()
460 /* vm_validate_pt_pd_bos - Validate page table and directory BOs
469 struct amdgpu_bo *pd = vm->root.bo; in vm_validate_pt_pd_bos()
470 struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev); in vm_validate_pt_pd_bos()
479 vm->pd_phys_addr = amdgpu_gmc_pd_addr(vm->root.bo); in vm_validate_pt_pd_bos()
486 struct amdgpu_bo *pd = vm->root.bo; in vm_update_pds()
487 struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev); in vm_update_pds()
494 return amdgpu_sync_fence(sync, vm->last_update); in vm_update_pds()
497 static uint64_t get_pte_flags(struct amdgpu_device *adev, struct kgd_mem *mem) in get_pte_flags() argument
502 if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE) in get_pte_flags()
504 if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE) in get_pte_flags()
511 * create_sg_table() - Create an sg_table for a contiguous DMA addr range
534 sg_dma_address(sg->sgl) = addr; in create_sg_table()
535 sg->sgl->length = size; in create_sg_table()
537 sg->sgl->dma_length = size; in create_sg_table()
543 kfd_mem_dmamap_userptr(struct kgd_mem *mem, in kfd_mem_dmamap_userptr() argument
547 mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ? in kfd_mem_dmamap_userptr()
550 struct amdgpu_bo *bo = attachment->bo_va->base.bo; in kfd_mem_dmamap_userptr()
551 struct amdgpu_device *adev = attachment->adev; in kfd_mem_dmamap_userptr()
552 struct ttm_tt *src_ttm = mem->bo->tbo.ttm; in kfd_mem_dmamap_userptr()
553 struct ttm_tt *ttm = bo->tbo.ttm; in kfd_mem_dmamap_userptr()
556 if (WARN_ON(ttm->num_pages != src_ttm->num_pages)) in kfd_mem_dmamap_userptr()
557 return -EINVAL; in kfd_mem_dmamap_userptr()
559 ttm->sg = kmalloc(sizeof(*ttm->sg), GFP_KERNEL); in kfd_mem_dmamap_userptr()
560 if (unlikely(!ttm->sg)) in kfd_mem_dmamap_userptr()
561 return -ENOMEM; in kfd_mem_dmamap_userptr()
564 ret = sg_alloc_table_from_pages(ttm->sg, src_ttm->pages, in kfd_mem_dmamap_userptr()
565 ttm->num_pages, 0, in kfd_mem_dmamap_userptr()
566 (u64)ttm->num_pages << PAGE_SHIFT, in kfd_mem_dmamap_userptr()
571 ret = dma_map_sgtable(adev->dev, ttm->sg, direction, 0); in kfd_mem_dmamap_userptr()
576 ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); in kfd_mem_dmamap_userptr()
583 dma_unmap_sgtable(adev->dev, ttm->sg, direction, 0); in kfd_mem_dmamap_userptr()
586 sg_free_table(ttm->sg); in kfd_mem_dmamap_userptr()
588 kfree(ttm->sg); in kfd_mem_dmamap_userptr()
589 ttm->sg = NULL; in kfd_mem_dmamap_userptr()
597 struct amdgpu_bo *bo = attachment->bo_va->base.bo; in kfd_mem_dmamap_dmabuf()
601 ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); in kfd_mem_dmamap_dmabuf()
606 return ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); in kfd_mem_dmamap_dmabuf()
610 * kfd_mem_dmamap_sg_bo() - Create DMA mapped sg_table to access DOORBELL or MMIO BO
611 * @mem: SG BO of the DOORBELL or MMIO resource on the owning device
623 * - Obtain DMA mapped address of DOORBELL or MMIO memory that could be used
625 * - Signal TTM to mark memory pointed to by requesting device's BO as GPU
630 * - Mapping of DOORBELL or MMIO BO of same or peer device
631 * - Validating an evicted DOOREBELL or MMIO BO on device seeking access
633 * Return: ZERO if successful, NON-ZERO otherwise
636 kfd_mem_dmamap_sg_bo(struct kgd_mem *mem, in kfd_mem_dmamap_sg_bo() argument
640 struct amdgpu_bo *bo = attachment->bo_va->base.bo; in kfd_mem_dmamap_sg_bo()
641 struct amdgpu_device *adev = attachment->adev; in kfd_mem_dmamap_sg_bo()
642 struct ttm_tt *ttm = bo->tbo.ttm; in kfd_mem_dmamap_sg_bo()
649 mmio = (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP); in kfd_mem_dmamap_sg_bo()
650 if (unlikely(ttm->sg)) { in kfd_mem_dmamap_sg_bo()
651 pr_err("SG Table of %d BO for peer device is UNEXPECTEDLY NON-NULL", mmio); in kfd_mem_dmamap_sg_bo()
652 return -EINVAL; in kfd_mem_dmamap_sg_bo()
655 dir = mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ? in kfd_mem_dmamap_sg_bo()
657 dma_addr = mem->bo->tbo.sg->sgl->dma_address; in kfd_mem_dmamap_sg_bo()
658 pr_debug("%d BO size: %d\n", mmio, mem->bo->tbo.sg->sgl->length); in kfd_mem_dmamap_sg_bo()
660 dma_addr = dma_map_resource(adev->dev, dma_addr, in kfd_mem_dmamap_sg_bo()
661 mem->bo->tbo.sg->sgl->length, dir, DMA_ATTR_SKIP_CPU_SYNC); in kfd_mem_dmamap_sg_bo()
662 ret = dma_mapping_error(adev->dev, dma_addr); in kfd_mem_dmamap_sg_bo()
667 ttm->sg = create_sg_table(dma_addr, mem->bo->tbo.sg->sgl->length); in kfd_mem_dmamap_sg_bo()
668 if (unlikely(!ttm->sg)) { in kfd_mem_dmamap_sg_bo()
669 ret = -ENOMEM; in kfd_mem_dmamap_sg_bo()
674 ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); in kfd_mem_dmamap_sg_bo()
681 sg_free_table(ttm->sg); in kfd_mem_dmamap_sg_bo()
682 kfree(ttm->sg); in kfd_mem_dmamap_sg_bo()
683 ttm->sg = NULL; in kfd_mem_dmamap_sg_bo()
685 dma_unmap_resource(adev->dev, dma_addr, mem->bo->tbo.sg->sgl->length, in kfd_mem_dmamap_sg_bo()
691 kfd_mem_dmamap_attachment(struct kgd_mem *mem, in kfd_mem_dmamap_attachment() argument
694 switch (attachment->type) { in kfd_mem_dmamap_attachment()
698 return kfd_mem_dmamap_userptr(mem, attachment); in kfd_mem_dmamap_attachment()
702 return kfd_mem_dmamap_sg_bo(mem, attachment); in kfd_mem_dmamap_attachment()
706 return -EINVAL; in kfd_mem_dmamap_attachment()
710 kfd_mem_dmaunmap_userptr(struct kgd_mem *mem, in kfd_mem_dmaunmap_userptr() argument
714 mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ? in kfd_mem_dmaunmap_userptr()
717 struct amdgpu_bo *bo = attachment->bo_va->base.bo; in kfd_mem_dmaunmap_userptr()
718 struct amdgpu_device *adev = attachment->adev; in kfd_mem_dmaunmap_userptr()
719 struct ttm_tt *ttm = bo->tbo.ttm; in kfd_mem_dmaunmap_userptr()
721 if (unlikely(!ttm->sg)) in kfd_mem_dmaunmap_userptr()
725 ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); in kfd_mem_dmaunmap_userptr()
727 dma_unmap_sgtable(adev->dev, ttm->sg, direction, 0); in kfd_mem_dmaunmap_userptr()
728 sg_free_table(ttm->sg); in kfd_mem_dmaunmap_userptr()
729 kfree(ttm->sg); in kfd_mem_dmaunmap_userptr()
730 ttm->sg = NULL; in kfd_mem_dmaunmap_userptr()
736 /* This is a no-op. We don't want to trigger eviction fences when in kfd_mem_dmaunmap_dmabuf()
743 * kfd_mem_dmaunmap_sg_bo() - Free DMA mapped sg_table of DOORBELL or MMIO BO
744 * @mem: SG BO of the DOORBELL or MMIO resource on the owning device
748 * - Signal TTM to mark memory pointed to by BO as GPU inaccessible
749 * - Free SG Table that is used to encapsulate DMA mapped memory of
759 kfd_mem_dmaunmap_sg_bo(struct kgd_mem *mem, in kfd_mem_dmaunmap_sg_bo() argument
763 struct amdgpu_bo *bo = attachment->bo_va->base.bo; in kfd_mem_dmaunmap_sg_bo()
764 struct amdgpu_device *adev = attachment->adev; in kfd_mem_dmaunmap_sg_bo()
765 struct ttm_tt *ttm = bo->tbo.ttm; in kfd_mem_dmaunmap_sg_bo()
768 if (unlikely(!ttm->sg)) { in kfd_mem_dmaunmap_sg_bo()
774 ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); in kfd_mem_dmaunmap_sg_bo()
776 dir = mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ? in kfd_mem_dmaunmap_sg_bo()
778 dma_unmap_resource(adev->dev, ttm->sg->sgl->dma_address, in kfd_mem_dmaunmap_sg_bo()
779 ttm->sg->sgl->length, dir, DMA_ATTR_SKIP_CPU_SYNC); in kfd_mem_dmaunmap_sg_bo()
780 sg_free_table(ttm->sg); in kfd_mem_dmaunmap_sg_bo()
781 kfree(ttm->sg); in kfd_mem_dmaunmap_sg_bo()
782 ttm->sg = NULL; in kfd_mem_dmaunmap_sg_bo()
783 bo->tbo.sg = NULL; in kfd_mem_dmaunmap_sg_bo()
787 kfd_mem_dmaunmap_attachment(struct kgd_mem *mem, in kfd_mem_dmaunmap_attachment() argument
790 switch (attachment->type) { in kfd_mem_dmaunmap_attachment()
794 kfd_mem_dmaunmap_userptr(mem, attachment); in kfd_mem_dmaunmap_attachment()
800 kfd_mem_dmaunmap_sg_bo(mem, attachment); in kfd_mem_dmaunmap_attachment()
807 static int kfd_mem_export_dmabuf(struct kgd_mem *mem) in kfd_mem_export_dmabuf() argument
809 if (!mem->dmabuf) { in kfd_mem_export_dmabuf()
814 bo_adev = amdgpu_ttm_adev(mem->bo->tbo.bdev); in kfd_mem_export_dmabuf()
815 r = drm_gem_prime_handle_to_fd(&bo_adev->ddev, bo_adev->kfd.client.file, in kfd_mem_export_dmabuf()
816 mem->gem_handle, in kfd_mem_export_dmabuf()
817 mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ? in kfd_mem_export_dmabuf()
825 mem->dmabuf = dmabuf; in kfd_mem_export_dmabuf()
832 kfd_mem_attach_dmabuf(struct amdgpu_device *adev, struct kgd_mem *mem, in kfd_mem_attach_dmabuf() argument
838 ret = kfd_mem_export_dmabuf(mem); in kfd_mem_attach_dmabuf()
842 gobj = amdgpu_gem_prime_import(adev_to_drm(adev), mem->dmabuf); in kfd_mem_attach_dmabuf()
847 (*bo)->flags |= AMDGPU_GEM_CREATE_PREEMPTIBLE; in kfd_mem_attach_dmabuf()
852 /* kfd_mem_attach - Add a BO to a VM
861 * 3. Determine ASIC-specific PTE flags
865 static int kfd_mem_attach(struct amdgpu_device *adev, struct kgd_mem *mem, in kfd_mem_attach() argument
868 struct amdgpu_device *bo_adev = amdgpu_ttm_adev(mem->bo->tbo.bdev); in kfd_mem_attach()
869 unsigned long bo_size = mem->bo->tbo.base.size; in kfd_mem_attach()
870 uint64_t va = mem->va; in kfd_mem_attach()
879 return -EINVAL; in kfd_mem_attach()
890 if ((adev != bo_adev && !adev->gmc.is_app_apu) && in kfd_mem_attach()
891 ((mem->domain == AMDGPU_GEM_DOMAIN_VRAM) || in kfd_mem_attach()
892 (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) || in kfd_mem_attach()
893 (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP))) { in kfd_mem_attach()
894 if (mem->domain == AMDGPU_GEM_DOMAIN_VRAM) in kfd_mem_attach()
897 return -EINVAL; in kfd_mem_attach()
903 ret = -ENOMEM; in kfd_mem_attach()
907 pr_debug("\t add VA 0x%llx - 0x%llx to vm %p\n", va, in kfd_mem_attach()
910 if ((adev == bo_adev && !(mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) || in kfd_mem_attach()
911 (amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm) && reuse_dmamap(adev, bo_adev)) || in kfd_mem_attach()
912 (mem->domain == AMDGPU_GEM_DOMAIN_GTT && reuse_dmamap(adev, bo_adev)) || in kfd_mem_attach()
918 attachment[i]->type = KFD_MEM_ATT_SHARED; in kfd_mem_attach()
919 bo[i] = mem->bo; in kfd_mem_attach()
920 drm_gem_object_get(&bo[i]->tbo.base); in kfd_mem_attach()
923 attachment[i]->type = KFD_MEM_ATT_SHARED; in kfd_mem_attach()
925 drm_gem_object_get(&bo[i]->tbo.base); in kfd_mem_attach()
926 } else if (amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm)) { in kfd_mem_attach()
927 /* Create an SG BO to DMA-map userptrs on other GPUs */ in kfd_mem_attach()
928 attachment[i]->type = KFD_MEM_ATT_USERPTR; in kfd_mem_attach()
929 ret = create_dmamap_sg_bo(adev, mem, &bo[i]); in kfd_mem_attach()
933 } else if (mem->bo->tbo.type == ttm_bo_type_sg) { in kfd_mem_attach()
934 WARN_ONCE(!(mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL || in kfd_mem_attach()
935 mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP), in kfd_mem_attach()
937 attachment[i]->type = KFD_MEM_ATT_SG; in kfd_mem_attach()
938 ret = create_dmamap_sg_bo(adev, mem, &bo[i]); in kfd_mem_attach()
942 } else if (mem->domain == AMDGPU_GEM_DOMAIN_GTT || in kfd_mem_attach()
943 mem->domain == AMDGPU_GEM_DOMAIN_VRAM) { in kfd_mem_attach()
944 attachment[i]->type = KFD_MEM_ATT_DMABUF; in kfd_mem_attach()
945 ret = kfd_mem_attach_dmabuf(adev, mem, &bo[i]); in kfd_mem_attach()
951 ret = -EINVAL; in kfd_mem_attach()
965 ++bo_va->ref_count; in kfd_mem_attach()
966 attachment[i]->bo_va = bo_va; in kfd_mem_attach()
968 if (unlikely(!attachment[i]->bo_va)) { in kfd_mem_attach()
969 ret = -ENOMEM; in kfd_mem_attach()
974 attachment[i]->va = va; in kfd_mem_attach()
975 attachment[i]->pte_flags = get_pte_flags(adev, mem); in kfd_mem_attach()
976 attachment[i]->adev = adev; in kfd_mem_attach()
977 list_add(&attachment[i]->list, &mem->attachments); in kfd_mem_attach()
985 for (; i >= 0; i--) { in kfd_mem_attach()
988 if (attachment[i]->bo_va) { in kfd_mem_attach()
990 if (--attachment[i]->bo_va->ref_count == 0) in kfd_mem_attach()
991 amdgpu_vm_bo_del(adev, attachment[i]->bo_va); in kfd_mem_attach()
993 list_del(&attachment[i]->list); in kfd_mem_attach()
996 drm_gem_object_put(&bo[i]->tbo.base); in kfd_mem_attach()
1004 struct amdgpu_bo *bo = attachment->bo_va->base.bo; in kfd_mem_detach()
1007 attachment->va, attachment); in kfd_mem_detach()
1008 if (--attachment->bo_va->ref_count == 0) in kfd_mem_detach()
1009 amdgpu_vm_bo_del(attachment->adev, attachment->bo_va); in kfd_mem_detach()
1010 drm_gem_object_put(&bo->tbo.base); in kfd_mem_detach()
1011 list_del(&attachment->list); in kfd_mem_detach()
1015 static void add_kgd_mem_to_kfd_bo_list(struct kgd_mem *mem, in add_kgd_mem_to_kfd_bo_list() argument
1019 mutex_lock(&process_info->lock); in add_kgd_mem_to_kfd_bo_list()
1021 list_add_tail(&mem->validate_list, in add_kgd_mem_to_kfd_bo_list()
1022 &process_info->userptr_valid_list); in add_kgd_mem_to_kfd_bo_list()
1024 list_add_tail(&mem->validate_list, &process_info->kfd_bo_list); in add_kgd_mem_to_kfd_bo_list()
1025 mutex_unlock(&process_info->lock); in add_kgd_mem_to_kfd_bo_list()
1028 static void remove_kgd_mem_from_kfd_bo_list(struct kgd_mem *mem, in remove_kgd_mem_from_kfd_bo_list() argument
1031 mutex_lock(&process_info->lock); in remove_kgd_mem_from_kfd_bo_list()
1032 list_del(&mem->validate_list); in remove_kgd_mem_from_kfd_bo_list()
1033 mutex_unlock(&process_info->lock); in remove_kgd_mem_from_kfd_bo_list()
1043 * Takes the process_info->lock to protect against concurrent restore
1048 static int init_user_pages(struct kgd_mem *mem, uint64_t user_addr, in init_user_pages() argument
1051 struct amdkfd_process_info *process_info = mem->process_info; in init_user_pages()
1052 struct amdgpu_bo *bo = mem->bo; in init_user_pages()
1057 mutex_lock(&process_info->lock); in init_user_pages()
1059 ret = amdgpu_ttm_tt_set_userptr(&bo->tbo, user_addr, 0); in init_user_pages()
1079 mutex_lock(&process_info->notifier_lock); in init_user_pages()
1080 mem->invalid++; in init_user_pages()
1081 mutex_unlock(&process_info->notifier_lock); in init_user_pages()
1082 mutex_unlock(&process_info->lock); in init_user_pages()
1086 ret = amdgpu_ttm_tt_get_user_pages(bo, bo->tbo.ttm->pages, &range); in init_user_pages()
1097 amdgpu_bo_placement_from_domain(bo, mem->domain); in init_user_pages()
1098 ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); in init_user_pages()
1104 amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm, range); in init_user_pages()
1109 mutex_unlock(&process_info->lock); in init_user_pages()
1134 * reserve_bo_and_vm - reserve a BO and a VM unconditionally.
1135 * @mem: KFD BO structure.
1139 static int reserve_bo_and_vm(struct kgd_mem *mem, in reserve_bo_and_vm() argument
1143 struct amdgpu_bo *bo = mem->bo; in reserve_bo_and_vm()
1148 ctx->n_vms = 1; in reserve_bo_and_vm()
1149 ctx->sync = &mem->sync; in reserve_bo_and_vm()
1150 drm_exec_init(&ctx->exec, DRM_EXEC_INTERRUPTIBLE_WAIT, 0); in reserve_bo_and_vm()
1151 drm_exec_until_all_locked(&ctx->exec) { in reserve_bo_and_vm()
1152 ret = amdgpu_vm_lock_pd(vm, &ctx->exec, 2); in reserve_bo_and_vm()
1153 drm_exec_retry_on_contention(&ctx->exec); in reserve_bo_and_vm()
1157 ret = drm_exec_prepare_obj(&ctx->exec, &bo->tbo.base, 1); in reserve_bo_and_vm()
1158 drm_exec_retry_on_contention(&ctx->exec); in reserve_bo_and_vm()
1166 drm_exec_fini(&ctx->exec); in reserve_bo_and_vm()
1171 * reserve_bo_and_cond_vms - reserve a BO and some VMs conditionally
1172 * @mem: KFD BO structure.
1180 static int reserve_bo_and_cond_vms(struct kgd_mem *mem, in reserve_bo_and_cond_vms() argument
1185 struct amdgpu_bo *bo = mem->bo; in reserve_bo_and_cond_vms()
1188 ctx->sync = &mem->sync; in reserve_bo_and_cond_vms()
1189 drm_exec_init(&ctx->exec, DRM_EXEC_INTERRUPTIBLE_WAIT, 0); in reserve_bo_and_cond_vms()
1190 drm_exec_until_all_locked(&ctx->exec) { in reserve_bo_and_cond_vms()
1191 ctx->n_vms = 0; in reserve_bo_and_cond_vms()
1192 list_for_each_entry(entry, &mem->attachments, list) { in reserve_bo_and_cond_vms()
1193 if ((vm && vm != entry->bo_va->base.vm) || in reserve_bo_and_cond_vms()
1194 (entry->is_mapped != map_type in reserve_bo_and_cond_vms()
1198 ret = amdgpu_vm_lock_pd(entry->bo_va->base.vm, in reserve_bo_and_cond_vms()
1199 &ctx->exec, 2); in reserve_bo_and_cond_vms()
1200 drm_exec_retry_on_contention(&ctx->exec); in reserve_bo_and_cond_vms()
1203 ++ctx->n_vms; in reserve_bo_and_cond_vms()
1206 ret = drm_exec_prepare_obj(&ctx->exec, &bo->tbo.base, 1); in reserve_bo_and_cond_vms()
1207 drm_exec_retry_on_contention(&ctx->exec); in reserve_bo_and_cond_vms()
1215 drm_exec_fini(&ctx->exec); in reserve_bo_and_cond_vms()
1220 * unreserve_bo_and_vms - Unreserve BO and VMs from a reservation context
1235 ret = amdgpu_sync_wait(ctx->sync, intr); in unreserve_bo_and_vms()
1237 drm_exec_fini(&ctx->exec); in unreserve_bo_and_vms()
1238 ctx->sync = NULL; in unreserve_bo_and_vms()
1242 static void unmap_bo_from_gpuvm(struct kgd_mem *mem, in unmap_bo_from_gpuvm() argument
1246 struct amdgpu_bo_va *bo_va = entry->bo_va; in unmap_bo_from_gpuvm()
1247 struct amdgpu_device *adev = entry->adev; in unmap_bo_from_gpuvm()
1248 struct amdgpu_vm *vm = bo_va->base.vm; in unmap_bo_from_gpuvm()
1250 amdgpu_vm_bo_unmap(adev, bo_va, entry->va); in unmap_bo_from_gpuvm()
1252 amdgpu_vm_clear_freed(adev, vm, &bo_va->last_pt_update); in unmap_bo_from_gpuvm()
1254 amdgpu_sync_fence(sync, bo_va->last_pt_update); in unmap_bo_from_gpuvm()
1257 static int update_gpuvm_pte(struct kgd_mem *mem, in update_gpuvm_pte() argument
1261 struct amdgpu_bo_va *bo_va = entry->bo_va; in update_gpuvm_pte()
1262 struct amdgpu_device *adev = entry->adev; in update_gpuvm_pte()
1265 ret = kfd_mem_dmamap_attachment(mem, entry); in update_gpuvm_pte()
1276 return amdgpu_sync_fence(sync, bo_va->last_pt_update); in update_gpuvm_pte()
1279 static int map_bo_to_gpuvm(struct kgd_mem *mem, in map_bo_to_gpuvm() argument
1287 ret = amdgpu_vm_bo_map(entry->adev, entry->bo_va, entry->va, 0, in map_bo_to_gpuvm()
1288 amdgpu_bo_size(entry->bo_va->base.bo), in map_bo_to_gpuvm()
1289 entry->pte_flags); in map_bo_to_gpuvm()
1292 entry->va, ret); in map_bo_to_gpuvm()
1299 ret = update_gpuvm_pte(mem, entry, sync); in map_bo_to_gpuvm()
1308 unmap_bo_from_gpuvm(mem, entry, sync); in map_bo_to_gpuvm()
1309 kfd_mem_dmaunmap_attachment(mem, entry); in map_bo_to_gpuvm()
1318 list_for_each_entry(peer_vm, &process_info->vm_list_head, in process_validate_vms()
1334 list_for_each_entry(peer_vm, &process_info->vm_list_head, in process_sync_pds_resv()
1336 struct amdgpu_bo *pd = peer_vm->root.bo; in process_sync_pds_resv()
1338 ret = amdgpu_sync_resv(NULL, sync, pd->tbo.base.resv, in process_sync_pds_resv()
1354 list_for_each_entry(peer_vm, &process_info->vm_list_head, in process_update_pds()
1373 return -ENOMEM; in init_kfd_vm()
1375 mutex_init(&info->lock); in init_kfd_vm()
1376 mutex_init(&info->notifier_lock); in init_kfd_vm()
1377 INIT_LIST_HEAD(&info->vm_list_head); in init_kfd_vm()
1378 INIT_LIST_HEAD(&info->kfd_bo_list); in init_kfd_vm()
1379 INIT_LIST_HEAD(&info->userptr_valid_list); in init_kfd_vm()
1380 INIT_LIST_HEAD(&info->userptr_inval_list); in init_kfd_vm()
1382 info->eviction_fence = in init_kfd_vm()
1384 current->mm, in init_kfd_vm()
1386 if (!info->eviction_fence) { in init_kfd_vm()
1388 ret = -ENOMEM; in init_kfd_vm()
1392 info->pid = get_task_pid(current->group_leader, PIDTYPE_PID); in init_kfd_vm()
1393 INIT_DELAYED_WORK(&info->restore_userptr_work, in init_kfd_vm()
1399 vm->process_info = *process_info; in init_kfd_vm()
1402 ret = amdgpu_bo_reserve(vm->root.bo, true); in init_kfd_vm()
1410 ret = amdgpu_bo_sync_wait(vm->root.bo, in init_kfd_vm()
1414 ret = dma_resv_reserve_fences(vm->root.bo->tbo.base.resv, 1); in init_kfd_vm()
1417 dma_resv_add_fence(vm->root.bo->tbo.base.resv, in init_kfd_vm()
1418 &vm->process_info->eviction_fence->base, in init_kfd_vm()
1420 amdgpu_bo_unreserve(vm->root.bo); in init_kfd_vm()
1423 mutex_lock(&vm->process_info->lock); in init_kfd_vm()
1424 list_add_tail(&vm->vm_list_node, in init_kfd_vm()
1425 &(vm->process_info->vm_list_head)); in init_kfd_vm()
1426 vm->process_info->n_vms++; in init_kfd_vm()
1428 *ef = dma_fence_get(&vm->process_info->eviction_fence->base); in init_kfd_vm()
1429 mutex_unlock(&vm->process_info->lock); in init_kfd_vm()
1436 amdgpu_bo_unreserve(vm->root.bo); in init_kfd_vm()
1438 vm->process_info = NULL; in init_kfd_vm()
1440 dma_fence_put(&info->eviction_fence->base); in init_kfd_vm()
1442 put_pid(info->pid); in init_kfd_vm()
1444 mutex_destroy(&info->lock); in init_kfd_vm()
1445 mutex_destroy(&info->notifier_lock); in init_kfd_vm()
1452 * amdgpu_amdkfd_gpuvm_pin_bo() - Pins a BO using following criteria
1456 * - USERPTR BOs are UNPINNABLE and will return error
1457 * - All other BO types (GTT, VRAM, MMIO and DOORBELL) will have their
1460 * Return: ZERO if successful in pinning, Non-Zero in case of error.
1481 * amdgpu_amdkfd_gpuvm_unpin_bo() - Unpins BO using following criteria
1484 * - Is a illegal request for USERPTR BOs and is ignored
1485 * - All other BO types (GTT, VRAM, MMIO and DOORBELL) will have their
1509 if (avm->pasid) { in amdgpu_amdkfd_gpuvm_set_vm_pasid()
1510 amdgpu_pasid_free(avm->pasid); in amdgpu_amdkfd_gpuvm_set_vm_pasid()
1529 if (avm->process_info) in amdgpu_amdkfd_gpuvm_acquire_process_vm()
1530 return -EINVAL; in amdgpu_amdkfd_gpuvm_acquire_process_vm()
1550 struct amdkfd_process_info *process_info = vm->process_info; in amdgpu_amdkfd_gpuvm_destroy_cb()
1556 mutex_lock(&process_info->lock); in amdgpu_amdkfd_gpuvm_destroy_cb()
1557 process_info->n_vms--; in amdgpu_amdkfd_gpuvm_destroy_cb()
1558 list_del(&vm->vm_list_node); in amdgpu_amdkfd_gpuvm_destroy_cb()
1559 mutex_unlock(&process_info->lock); in amdgpu_amdkfd_gpuvm_destroy_cb()
1561 vm->process_info = NULL; in amdgpu_amdkfd_gpuvm_destroy_cb()
1563 /* Release per-process resources when last compute VM is destroyed */ in amdgpu_amdkfd_gpuvm_destroy_cb()
1564 if (!process_info->n_vms) { in amdgpu_amdkfd_gpuvm_destroy_cb()
1565 WARN_ON(!list_empty(&process_info->kfd_bo_list)); in amdgpu_amdkfd_gpuvm_destroy_cb()
1566 WARN_ON(!list_empty(&process_info->userptr_valid_list)); in amdgpu_amdkfd_gpuvm_destroy_cb()
1567 WARN_ON(!list_empty(&process_info->userptr_inval_list)); in amdgpu_amdkfd_gpuvm_destroy_cb()
1569 dma_fence_put(&process_info->eviction_fence->base); in amdgpu_amdkfd_gpuvm_destroy_cb()
1570 cancel_delayed_work_sync(&process_info->restore_userptr_work); in amdgpu_amdkfd_gpuvm_destroy_cb()
1571 put_pid(process_info->pid); in amdgpu_amdkfd_gpuvm_destroy_cb()
1572 mutex_destroy(&process_info->lock); in amdgpu_amdkfd_gpuvm_destroy_cb()
1573 mutex_destroy(&process_info->notifier_lock); in amdgpu_amdkfd_gpuvm_destroy_cb()
1602 struct amdgpu_bo *pd = avm->root.bo; in amdgpu_amdkfd_gpuvm_get_process_page_dir()
1603 struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev); in amdgpu_amdkfd_gpuvm_get_process_page_dir()
1605 if (adev->asic_type < CHIP_VEGA10) in amdgpu_amdkfd_gpuvm_get_process_page_dir()
1606 return avm->pd_phys_addr >> AMDGPU_GPU_PAGE_SHIFT; in amdgpu_amdkfd_gpuvm_get_process_page_dir()
1607 return avm->pd_phys_addr; in amdgpu_amdkfd_gpuvm_get_process_page_dir()
1614 mutex_lock(&pinfo->lock); in amdgpu_amdkfd_block_mmu_notifications()
1615 WRITE_ONCE(pinfo->block_mmu_notifications, true); in amdgpu_amdkfd_block_mmu_notifications()
1616 mutex_unlock(&pinfo->lock); in amdgpu_amdkfd_block_mmu_notifications()
1624 mutex_lock(&pinfo->lock); in amdgpu_amdkfd_criu_resume()
1626 mutex_lock(&pinfo->notifier_lock); in amdgpu_amdkfd_criu_resume()
1627 pinfo->evicted_bos++; in amdgpu_amdkfd_criu_resume()
1628 mutex_unlock(&pinfo->notifier_lock); in amdgpu_amdkfd_criu_resume()
1629 if (!READ_ONCE(pinfo->block_mmu_notifications)) { in amdgpu_amdkfd_criu_resume()
1630 ret = -EINVAL; in amdgpu_amdkfd_criu_resume()
1633 WRITE_ONCE(pinfo->block_mmu_notifications, false); in amdgpu_amdkfd_criu_resume()
1635 &pinfo->restore_userptr_work, 0); in amdgpu_amdkfd_criu_resume()
1638 mutex_unlock(&pinfo->lock); in amdgpu_amdkfd_criu_resume()
1652 - adev->kfd.vram_used_aligned[xcp_id] in amdgpu_amdkfd_get_available_memory()
1653 - atomic64_read(&adev->vram_pin_size) in amdgpu_amdkfd_get_available_memory()
1654 - reserved_for_pt; in amdgpu_amdkfd_get_available_memory()
1656 if (adev->gmc.is_app_apu) { in amdgpu_amdkfd_get_available_memory()
1659 kfd_mem_limit.max_system_mem_limit - in amdgpu_amdkfd_get_available_memory()
1662 ttm_mem_available = kfd_mem_limit.max_ttm_mem_limit - in amdgpu_amdkfd_get_available_memory()
1682 void *drm_priv, struct kgd_mem **mem, in amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu() argument
1694 int8_t xcp_id = -1; in amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu()
1704 if (adev->gmc.is_app_apu) { in amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu()
1713 xcp_id = fpriv->xcp_id == AMDGPU_XCP_NO_PARTITION ? in amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu()
1714 0 : fpriv->xcp_id; in amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu()
1725 return -EINVAL; in amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu()
1731 return -EINVAL; in amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu()
1734 return -ENOMEM; in amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu()
1736 return -EINVAL; in amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu()
1747 *mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL); in amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu()
1748 if (!*mem) { in amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu()
1749 ret = -ENOMEM; in amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu()
1752 INIT_LIST_HEAD(&(*mem)->attachments); in amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu()
1753 mutex_init(&(*mem)->lock); in amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu()
1754 (*mem)->aql_queue = !!(flags & KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM); in amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu()
1760 if ((*mem)->aql_queue) in amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu()
1764 (*mem)->alloc_flags = flags; in amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu()
1766 amdgpu_sync_create(&(*mem)->sync); in amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu()
1776 va, (*mem)->aql_queue ? size << 1 : size, in amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu()
1786 ret = drm_vma_node_allow(&gobj->vma_node, drm_priv); in amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu()
1791 ret = drm_gem_handle_create(adev->kfd.client.file, gobj, &(*mem)->gem_handle); in amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu()
1796 bo->tbo.sg = sg; in amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu()
1797 bo->tbo.ttm->sg = sg; in amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu()
1799 bo->kfd_bo = *mem; in amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu()
1800 (*mem)->bo = bo; in amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu()
1802 bo->flags |= AMDGPU_AMDKFD_CREATE_USERPTR_BO; in amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu()
1804 (*mem)->va = va; in amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu()
1805 (*mem)->domain = domain; in amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu()
1806 (*mem)->mapped_to_gpu_memory = 0; in amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu()
1807 (*mem)->process_info = avm->process_info; in amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu()
1809 add_kgd_mem_to_kfd_bo_list(*mem, avm->process_info, user_addr); in amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu()
1813 ret = init_user_pages(*mem, user_addr, criu_resume); in amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu()
1823 bo->allowed_domains = AMDGPU_GEM_DOMAIN_GTT; in amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu()
1824 bo->preferred_domains = AMDGPU_GEM_DOMAIN_GTT; in amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu()
1826 mutex_lock(&avm->process_info->lock); in amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu()
1827 if (avm->process_info->eviction_fence && in amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu()
1828 !dma_fence_is_signaled(&avm->process_info->eviction_fence->base)) in amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu()
1830 &avm->process_info->eviction_fence->base); in amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu()
1831 mutex_unlock(&avm->process_info->lock); in amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu()
1844 remove_kgd_mem_from_kfd_bo_list(*mem, avm->process_info); in amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu()
1845 drm_gem_handle_delete(adev->kfd.client.file, (*mem)->gem_handle); in amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu()
1847 drm_vma_node_revoke(&gobj->vma_node, drm_priv); in amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu()
1849 /* Don't unreserve system mem limit twice */ in amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu()
1854 mutex_destroy(&(*mem)->lock); in amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu()
1858 kfree(*mem); in amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu()
1868 struct amdgpu_device *adev, struct kgd_mem *mem, void *drm_priv, in amdgpu_amdkfd_gpuvm_free_memory_of_gpu() argument
1871 struct amdkfd_process_info *process_info = mem->process_info; in amdgpu_amdkfd_gpuvm_free_memory_of_gpu()
1872 unsigned long bo_size = mem->bo->tbo.base.size; in amdgpu_amdkfd_gpuvm_free_memory_of_gpu()
1873 bool use_release_notifier = (mem->bo->kfd_bo == mem); in amdgpu_amdkfd_gpuvm_free_memory_of_gpu()
1880 mutex_lock(&mem->lock); in amdgpu_amdkfd_gpuvm_free_memory_of_gpu()
1883 if (mem->alloc_flags & in amdgpu_amdkfd_gpuvm_free_memory_of_gpu()
1886 amdgpu_amdkfd_gpuvm_unpin_bo(mem->bo); in amdgpu_amdkfd_gpuvm_free_memory_of_gpu()
1889 mapped_to_gpu_memory = mem->mapped_to_gpu_memory; in amdgpu_amdkfd_gpuvm_free_memory_of_gpu()
1890 is_imported = mem->is_imported; in amdgpu_amdkfd_gpuvm_free_memory_of_gpu()
1891 mutex_unlock(&mem->lock); in amdgpu_amdkfd_gpuvm_free_memory_of_gpu()
1892 /* lock is not needed after this, since mem is unused and will in amdgpu_amdkfd_gpuvm_free_memory_of_gpu()
1898 mem->va, bo_size); in amdgpu_amdkfd_gpuvm_free_memory_of_gpu()
1899 return -EBUSY; in amdgpu_amdkfd_gpuvm_free_memory_of_gpu()
1903 mutex_lock(&process_info->lock); in amdgpu_amdkfd_gpuvm_free_memory_of_gpu()
1904 list_del(&mem->validate_list); in amdgpu_amdkfd_gpuvm_free_memory_of_gpu()
1905 mutex_unlock(&process_info->lock); in amdgpu_amdkfd_gpuvm_free_memory_of_gpu()
1908 if (amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm)) { in amdgpu_amdkfd_gpuvm_free_memory_of_gpu()
1909 amdgpu_hmm_unregister(mem->bo); in amdgpu_amdkfd_gpuvm_free_memory_of_gpu()
1910 mutex_lock(&process_info->notifier_lock); in amdgpu_amdkfd_gpuvm_free_memory_of_gpu()
1911 amdgpu_ttm_tt_discard_user_pages(mem->bo->tbo.ttm, mem->range); in amdgpu_amdkfd_gpuvm_free_memory_of_gpu()
1912 mutex_unlock(&process_info->notifier_lock); in amdgpu_amdkfd_gpuvm_free_memory_of_gpu()
1915 ret = reserve_bo_and_cond_vms(mem, NULL, BO_VM_ALL, &ctx); in amdgpu_amdkfd_gpuvm_free_memory_of_gpu()
1919 amdgpu_amdkfd_remove_eviction_fence(mem->bo, in amdgpu_amdkfd_gpuvm_free_memory_of_gpu()
1920 process_info->eviction_fence); in amdgpu_amdkfd_gpuvm_free_memory_of_gpu()
1921 pr_debug("Release VA 0x%llx - 0x%llx\n", mem->va, in amdgpu_amdkfd_gpuvm_free_memory_of_gpu()
1922 mem->va + bo_size * (1 + mem->aql_queue)); in amdgpu_amdkfd_gpuvm_free_memory_of_gpu()
1925 list_for_each_entry_safe(entry, tmp, &mem->attachments, list) { in amdgpu_amdkfd_gpuvm_free_memory_of_gpu()
1926 kfd_mem_dmaunmap_attachment(mem, entry); in amdgpu_amdkfd_gpuvm_free_memory_of_gpu()
1933 amdgpu_sync_free(&mem->sync); in amdgpu_amdkfd_gpuvm_free_memory_of_gpu()
1938 if (mem->bo->tbo.sg) { in amdgpu_amdkfd_gpuvm_free_memory_of_gpu()
1939 sg_free_table(mem->bo->tbo.sg); in amdgpu_amdkfd_gpuvm_free_memory_of_gpu()
1940 kfree(mem->bo->tbo.sg); in amdgpu_amdkfd_gpuvm_free_memory_of_gpu()
1949 (mem->bo->preferred_domains == AMDGPU_GEM_DOMAIN_VRAM || in amdgpu_amdkfd_gpuvm_free_memory_of_gpu()
1950 (adev->gmc.is_app_apu && in amdgpu_amdkfd_gpuvm_free_memory_of_gpu()
1951 mem->bo->preferred_domains == AMDGPU_GEM_DOMAIN_GTT))) in amdgpu_amdkfd_gpuvm_free_memory_of_gpu()
1958 drm_vma_node_revoke(&mem->bo->tbo.base.vma_node, drm_priv); in amdgpu_amdkfd_gpuvm_free_memory_of_gpu()
1959 drm_gem_handle_delete(adev->kfd.client.file, mem->gem_handle); in amdgpu_amdkfd_gpuvm_free_memory_of_gpu()
1960 if (mem->dmabuf) { in amdgpu_amdkfd_gpuvm_free_memory_of_gpu()
1961 dma_buf_put(mem->dmabuf); in amdgpu_amdkfd_gpuvm_free_memory_of_gpu()
1962 mem->dmabuf = NULL; in amdgpu_amdkfd_gpuvm_free_memory_of_gpu()
1964 mutex_destroy(&mem->lock); in amdgpu_amdkfd_gpuvm_free_memory_of_gpu()
1967 * amdgpu_amdkfd_release_notify and kfree the mem struct. That's why in amdgpu_amdkfd_gpuvm_free_memory_of_gpu()
1970 drm_gem_object_put(&mem->bo->tbo.base); in amdgpu_amdkfd_gpuvm_free_memory_of_gpu()
1977 kfree(mem); in amdgpu_amdkfd_gpuvm_free_memory_of_gpu()
1983 struct amdgpu_device *adev, struct kgd_mem *mem, in amdgpu_amdkfd_gpuvm_map_memory_to_gpu() argument
1995 bo = mem->bo; in amdgpu_amdkfd_gpuvm_map_memory_to_gpu()
1998 return -EINVAL; in amdgpu_amdkfd_gpuvm_map_memory_to_gpu()
2005 mutex_lock(&mem->process_info->lock); in amdgpu_amdkfd_gpuvm_map_memory_to_gpu()
2011 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) { in amdgpu_amdkfd_gpuvm_map_memory_to_gpu()
2012 mutex_lock(&mem->process_info->notifier_lock); in amdgpu_amdkfd_gpuvm_map_memory_to_gpu()
2013 is_invalid_userptr = !!mem->invalid; in amdgpu_amdkfd_gpuvm_map_memory_to_gpu()
2014 mutex_unlock(&mem->process_info->notifier_lock); in amdgpu_amdkfd_gpuvm_map_memory_to_gpu()
2017 mutex_lock(&mem->lock); in amdgpu_amdkfd_gpuvm_map_memory_to_gpu()
2019 domain = mem->domain; in amdgpu_amdkfd_gpuvm_map_memory_to_gpu()
2020 bo_size = bo->tbo.base.size; in amdgpu_amdkfd_gpuvm_map_memory_to_gpu()
2022 pr_debug("Map VA 0x%llx - 0x%llx to vm %p domain %s\n", in amdgpu_amdkfd_gpuvm_map_memory_to_gpu()
2023 mem->va, in amdgpu_amdkfd_gpuvm_map_memory_to_gpu()
2024 mem->va + bo_size * (1 + mem->aql_queue), in amdgpu_amdkfd_gpuvm_map_memory_to_gpu()
2027 if (!kfd_mem_is_attached(avm, mem)) { in amdgpu_amdkfd_gpuvm_map_memory_to_gpu()
2028 ret = kfd_mem_attach(adev, mem, avm, mem->aql_queue); in amdgpu_amdkfd_gpuvm_map_memory_to_gpu()
2033 ret = reserve_bo_and_vm(mem, avm, &ctx); in amdgpu_amdkfd_gpuvm_map_memory_to_gpu()
2042 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) && in amdgpu_amdkfd_gpuvm_map_memory_to_gpu()
2043 bo->tbo.resource->mem_type == TTM_PL_SYSTEM) in amdgpu_amdkfd_gpuvm_map_memory_to_gpu()
2050 list_for_each_entry(entry, &mem->attachments, list) { in amdgpu_amdkfd_gpuvm_map_memory_to_gpu()
2051 if (entry->bo_va->base.vm != avm || entry->is_mapped) in amdgpu_amdkfd_gpuvm_map_memory_to_gpu()
2054 pr_debug("\t map VA 0x%llx - 0x%llx in entry %p\n", in amdgpu_amdkfd_gpuvm_map_memory_to_gpu()
2055 entry->va, entry->va + bo_size, entry); in amdgpu_amdkfd_gpuvm_map_memory_to_gpu()
2057 ret = map_bo_to_gpuvm(mem, entry, ctx.sync, in amdgpu_amdkfd_gpuvm_map_memory_to_gpu()
2070 entry->is_mapped = true; in amdgpu_amdkfd_gpuvm_map_memory_to_gpu()
2071 mem->mapped_to_gpu_memory++; in amdgpu_amdkfd_gpuvm_map_memory_to_gpu()
2073 mem->mapped_to_gpu_memory); in amdgpu_amdkfd_gpuvm_map_memory_to_gpu()
2083 mutex_unlock(&mem->process_info->lock); in amdgpu_amdkfd_gpuvm_map_memory_to_gpu()
2084 mutex_unlock(&mem->lock); in amdgpu_amdkfd_gpuvm_map_memory_to_gpu()
2088 int amdgpu_amdkfd_gpuvm_dmaunmap_mem(struct kgd_mem *mem, void *drm_priv) in amdgpu_amdkfd_gpuvm_dmaunmap_mem() argument
2096 mutex_lock(&mem->lock); in amdgpu_amdkfd_gpuvm_dmaunmap_mem()
2098 ret = amdgpu_bo_reserve(mem->bo, true); in amdgpu_amdkfd_gpuvm_dmaunmap_mem()
2102 list_for_each_entry(entry, &mem->attachments, list) { in amdgpu_amdkfd_gpuvm_dmaunmap_mem()
2103 if (entry->bo_va->base.vm != vm) in amdgpu_amdkfd_gpuvm_dmaunmap_mem()
2105 if (entry->bo_va->base.bo->tbo.ttm && in amdgpu_amdkfd_gpuvm_dmaunmap_mem()
2106 !entry->bo_va->base.bo->tbo.ttm->sg) in amdgpu_amdkfd_gpuvm_dmaunmap_mem()
2109 kfd_mem_dmaunmap_attachment(mem, entry); in amdgpu_amdkfd_gpuvm_dmaunmap_mem()
2112 amdgpu_bo_unreserve(mem->bo); in amdgpu_amdkfd_gpuvm_dmaunmap_mem()
2114 mutex_unlock(&mem->lock); in amdgpu_amdkfd_gpuvm_dmaunmap_mem()
2120 struct amdgpu_device *adev, struct kgd_mem *mem, void *drm_priv) in amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu() argument
2123 unsigned long bo_size = mem->bo->tbo.base.size; in amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu()
2128 mutex_lock(&mem->lock); in amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu()
2130 ret = reserve_bo_and_cond_vms(mem, avm, BO_VM_MAPPED, &ctx); in amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu()
2135 ret = -EINVAL; in amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu()
2143 pr_debug("Unmap VA 0x%llx - 0x%llx from vm %p\n", in amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu()
2144 mem->va, in amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu()
2145 mem->va + bo_size * (1 + mem->aql_queue), in amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu()
2148 list_for_each_entry(entry, &mem->attachments, list) { in amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu()
2149 if (entry->bo_va->base.vm != avm || !entry->is_mapped) in amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu()
2152 pr_debug("\t unmap VA 0x%llx - 0x%llx from entry %p\n", in amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu()
2153 entry->va, entry->va + bo_size, entry); in amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu()
2155 unmap_bo_from_gpuvm(mem, entry, ctx.sync); in amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu()
2156 entry->is_mapped = false; in amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu()
2158 mem->mapped_to_gpu_memory--; in amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu()
2160 mem->mapped_to_gpu_memory); in amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu()
2166 mutex_unlock(&mem->lock); in amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu()
2171 struct amdgpu_device *adev, struct kgd_mem *mem, bool intr) in amdgpu_amdkfd_gpuvm_sync_memory() argument
2178 mutex_lock(&mem->lock); in amdgpu_amdkfd_gpuvm_sync_memory()
2179 amdgpu_sync_clone(&mem->sync, &sync); in amdgpu_amdkfd_gpuvm_sync_memory()
2180 mutex_unlock(&mem->lock); in amdgpu_amdkfd_gpuvm_sync_memory()
2188 * amdgpu_amdkfd_map_gtt_bo_to_gart - Map BO to GART and increment reference count
2211 ret = amdgpu_ttm_alloc_gart(&bo->tbo); in amdgpu_amdkfd_map_gtt_bo_to_gart()
2218 bo, bo->vm_bo->vm->process_info->eviction_fence); in amdgpu_amdkfd_map_gtt_bo_to_gart()
2235 /** amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel() - Map a GTT BO for kernel CPU access
2237 * @mem: Buffer object to be mapped for CPU access
2248 int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct kgd_mem *mem, in amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel() argument
2252 struct amdgpu_bo *bo = mem->bo; in amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel()
2254 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) { in amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel()
2256 return -EINVAL; in amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel()
2259 mutex_lock(&mem->process_info->lock); in amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel()
2280 bo, mem->process_info->eviction_fence); in amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel()
2287 mutex_unlock(&mem->process_info->lock); in amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel()
2295 mutex_unlock(&mem->process_info->lock); in amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel()
2300 /** amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel() - Unmap a GTT BO for kernel CPU access
2302 * @mem: Buffer object to be unmapped for CPU access
2308 void amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel(struct kgd_mem *mem) in amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel() argument
2310 struct amdgpu_bo *bo = mem->bo; in amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel()
2319 struct kfd_vm_fault_info *mem) in amdgpu_amdkfd_gpuvm_get_vm_fault_info() argument
2321 if (atomic_read(&adev->gmc.vm_fault_info_updated) == 1) { in amdgpu_amdkfd_gpuvm_get_vm_fault_info()
2322 *mem = *adev->gmc.vm_fault_info; in amdgpu_amdkfd_gpuvm_get_vm_fault_info()
2324 atomic_set(&adev->gmc.vm_fault_info_updated, 0); in amdgpu_amdkfd_gpuvm_get_vm_fault_info()
2333 struct kgd_mem **mem, uint64_t *size, in import_obj_create() argument
2341 if (!(bo->preferred_domains & (AMDGPU_GEM_DOMAIN_VRAM | in import_obj_create()
2344 return -EINVAL; in import_obj_create()
2346 *mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL); in import_obj_create()
2347 if (!*mem) in import_obj_create()
2348 return -ENOMEM; in import_obj_create()
2350 ret = drm_vma_node_allow(&obj->vma_node, drm_priv); in import_obj_create()
2360 INIT_LIST_HEAD(&(*mem)->attachments); in import_obj_create()
2361 mutex_init(&(*mem)->lock); in import_obj_create()
2363 (*mem)->alloc_flags = in import_obj_create()
2364 ((bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) ? in import_obj_create()
2370 (*mem)->dmabuf = dma_buf; in import_obj_create()
2371 (*mem)->bo = bo; in import_obj_create()
2372 (*mem)->va = va; in import_obj_create()
2373 (*mem)->domain = (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) && !adev->gmc.is_app_apu ? in import_obj_create()
2376 (*mem)->mapped_to_gpu_memory = 0; in import_obj_create()
2377 (*mem)->process_info = avm->process_info; in import_obj_create()
2378 add_kgd_mem_to_kfd_bo_list(*mem, avm->process_info, false); in import_obj_create()
2379 amdgpu_sync_create(&(*mem)->sync); in import_obj_create()
2380 (*mem)->is_imported = true; in import_obj_create()
2382 mutex_lock(&avm->process_info->lock); in import_obj_create()
2383 if (avm->process_info->eviction_fence && in import_obj_create()
2384 !dma_fence_is_signaled(&avm->process_info->eviction_fence->base)) in import_obj_create()
2385 ret = amdgpu_amdkfd_bo_validate_and_fence(bo, (*mem)->domain, in import_obj_create()
2386 &avm->process_info->eviction_fence->base); in import_obj_create()
2387 mutex_unlock(&avm->process_info->lock); in import_obj_create()
2394 remove_kgd_mem_from_kfd_bo_list(*mem, avm->process_info); in import_obj_create()
2395 drm_vma_node_revoke(&obj->vma_node, drm_priv); in import_obj_create()
2397 kfree(*mem); in import_obj_create()
2403 struct kgd_mem **mem, uint64_t *size, in amdgpu_amdkfd_gpuvm_import_dmabuf_fd() argument
2410 ret = drm_gem_prime_fd_to_handle(&adev->ddev, adev->kfd.client.file, fd, in amdgpu_amdkfd_gpuvm_import_dmabuf_fd()
2414 obj = drm_gem_object_lookup(adev->kfd.client.file, handle); in amdgpu_amdkfd_gpuvm_import_dmabuf_fd()
2416 ret = -EINVAL; in amdgpu_amdkfd_gpuvm_import_dmabuf_fd()
2420 ret = import_obj_create(adev, obj->dma_buf, obj, va, drm_priv, mem, size, in amdgpu_amdkfd_gpuvm_import_dmabuf_fd()
2425 (*mem)->gem_handle = handle; in amdgpu_amdkfd_gpuvm_import_dmabuf_fd()
2432 drm_gem_handle_delete(adev->kfd.client.file, handle); in amdgpu_amdkfd_gpuvm_import_dmabuf_fd()
2436 int amdgpu_amdkfd_gpuvm_export_dmabuf(struct kgd_mem *mem, in amdgpu_amdkfd_gpuvm_export_dmabuf() argument
2441 mutex_lock(&mem->lock); in amdgpu_amdkfd_gpuvm_export_dmabuf()
2442 ret = kfd_mem_export_dmabuf(mem); in amdgpu_amdkfd_gpuvm_export_dmabuf()
2446 get_dma_buf(mem->dmabuf); in amdgpu_amdkfd_gpuvm_export_dmabuf()
2447 *dma_buf = mem->dmabuf; in amdgpu_amdkfd_gpuvm_export_dmabuf()
2449 mutex_unlock(&mem->lock); in amdgpu_amdkfd_gpuvm_export_dmabuf()
2464 unsigned long cur_seq, struct kgd_mem *mem) in amdgpu_amdkfd_evict_userptr() argument
2466 struct amdkfd_process_info *process_info = mem->process_info; in amdgpu_amdkfd_evict_userptr()
2472 if (READ_ONCE(process_info->block_mmu_notifications)) in amdgpu_amdkfd_evict_userptr()
2475 mutex_lock(&process_info->notifier_lock); in amdgpu_amdkfd_evict_userptr()
2478 mem->invalid++; in amdgpu_amdkfd_evict_userptr()
2479 if (++process_info->evicted_bos == 1) { in amdgpu_amdkfd_evict_userptr()
2481 r = kgd2kfd_quiesce_mm(mni->mm, in amdgpu_amdkfd_evict_userptr()
2486 &process_info->restore_userptr_work, in amdgpu_amdkfd_evict_userptr()
2489 mutex_unlock(&process_info->notifier_lock); in amdgpu_amdkfd_evict_userptr()
2503 struct kgd_mem *mem, *tmp_mem; in update_invalid_user_pages() local
2509 mutex_lock(&process_info->notifier_lock); in update_invalid_user_pages()
2512 list_for_each_entry_safe(mem, tmp_mem, in update_invalid_user_pages()
2513 &process_info->userptr_valid_list, in update_invalid_user_pages()
2515 if (mem->invalid) in update_invalid_user_pages()
2516 list_move_tail(&mem->validate_list, in update_invalid_user_pages()
2517 &process_info->userptr_inval_list); in update_invalid_user_pages()
2520 list_for_each_entry(mem, &process_info->userptr_inval_list, in update_invalid_user_pages()
2522 invalid = mem->invalid; in update_invalid_user_pages()
2529 bo = mem->bo; in update_invalid_user_pages()
2531 amdgpu_ttm_tt_discard_user_pages(bo->tbo.ttm, mem->range); in update_invalid_user_pages()
2532 mem->range = NULL; in update_invalid_user_pages()
2537 mutex_unlock(&process_info->notifier_lock); in update_invalid_user_pages()
2542 if (bo->tbo.resource->mem_type != TTM_PL_SYSTEM) { in update_invalid_user_pages()
2544 return -EAGAIN; in update_invalid_user_pages()
2546 ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); in update_invalid_user_pages()
2551 return -EAGAIN; in update_invalid_user_pages()
2556 ret = amdgpu_ttm_tt_get_user_pages(bo, bo->tbo.ttm->pages, in update_invalid_user_pages()
2557 &mem->range); in update_invalid_user_pages()
2561 /* Return -EFAULT bad address error as success. It will in update_invalid_user_pages()
2566 * Return other error -EBUSY or -ENOMEM to retry restore in update_invalid_user_pages()
2568 if (ret != -EFAULT) in update_invalid_user_pages()
2574 mutex_lock(&process_info->notifier_lock); in update_invalid_user_pages()
2579 if (mem->invalid != invalid) { in update_invalid_user_pages()
2580 ret = -EAGAIN; in update_invalid_user_pages()
2583 /* set mem valid if mem has hmm range associated */ in update_invalid_user_pages()
2584 if (mem->range) in update_invalid_user_pages()
2585 mem->invalid = 0; in update_invalid_user_pages()
2589 mutex_unlock(&process_info->notifier_lock); in update_invalid_user_pages()
2606 struct kgd_mem *mem, *tmp_mem; in validate_invalid_user_pages() local
2616 list_for_each_entry(peer_vm, &process_info->vm_list_head, in validate_invalid_user_pages()
2625 list_for_each_entry(mem, &process_info->userptr_inval_list, in validate_invalid_user_pages()
2629 gobj = &mem->bo->tbo.base; in validate_invalid_user_pages()
2642 list_for_each_entry_safe(mem, tmp_mem, in validate_invalid_user_pages()
2643 &process_info->userptr_inval_list, in validate_invalid_user_pages()
2647 bo = mem->bo; in validate_invalid_user_pages()
2650 if (bo->tbo.ttm->pages[0]) { in validate_invalid_user_pages()
2651 amdgpu_bo_placement_from_domain(bo, mem->domain); in validate_invalid_user_pages()
2652 ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); in validate_invalid_user_pages()
2665 list_for_each_entry(attachment, &mem->attachments, list) { in validate_invalid_user_pages()
2666 if (!attachment->is_mapped) in validate_invalid_user_pages()
2669 kfd_mem_dmaunmap_attachment(mem, attachment); in validate_invalid_user_pages()
2670 ret = update_gpuvm_pte(mem, attachment, &sync); in validate_invalid_user_pages()
2674 mutex_lock(&process_info->notifier_lock); in validate_invalid_user_pages()
2675 mem->invalid++; in validate_invalid_user_pages()
2676 mutex_unlock(&process_info->notifier_lock); in validate_invalid_user_pages()
2699 struct kgd_mem *mem, *tmp_mem; in confirm_valid_user_pages_locked() local
2702 list_for_each_entry_safe(mem, tmp_mem, in confirm_valid_user_pages_locked()
2703 &process_info->userptr_inval_list, in confirm_valid_user_pages_locked()
2707 /* keep mem without hmm range at userptr_inval_list */ in confirm_valid_user_pages_locked()
2708 if (!mem->range) in confirm_valid_user_pages_locked()
2711 /* Only check mem with hmm range associated */ in confirm_valid_user_pages_locked()
2713 mem->bo->tbo.ttm, mem->range); in confirm_valid_user_pages_locked()
2715 mem->range = NULL; in confirm_valid_user_pages_locked()
2717 WARN(!mem->invalid, "Invalid BO not marked invalid"); in confirm_valid_user_pages_locked()
2718 ret = -EAGAIN; in confirm_valid_user_pages_locked()
2722 if (mem->invalid) { in confirm_valid_user_pages_locked()
2724 ret = -EAGAIN; in confirm_valid_user_pages_locked()
2728 list_move_tail(&mem->validate_list, in confirm_valid_user_pages_locked()
2729 &process_info->userptr_valid_list); in confirm_valid_user_pages_locked()
2751 mutex_lock(&process_info->notifier_lock); in amdgpu_amdkfd_restore_userptr_worker()
2752 evicted_bos = process_info->evicted_bos; in amdgpu_amdkfd_restore_userptr_worker()
2753 mutex_unlock(&process_info->notifier_lock); in amdgpu_amdkfd_restore_userptr_worker()
2758 usertask = get_pid_task(process_info->pid, PIDTYPE_PID); in amdgpu_amdkfd_restore_userptr_worker()
2767 mutex_lock(&process_info->lock); in amdgpu_amdkfd_restore_userptr_worker()
2775 if (!list_empty(&process_info->userptr_inval_list)) { in amdgpu_amdkfd_restore_userptr_worker()
2784 mutex_lock(&process_info->notifier_lock); in amdgpu_amdkfd_restore_userptr_worker()
2785 if (process_info->evicted_bos != evicted_bos) in amdgpu_amdkfd_restore_userptr_worker()
2793 process_info->evicted_bos = evicted_bos = 0; in amdgpu_amdkfd_restore_userptr_worker()
2803 mutex_unlock(&process_info->notifier_lock); in amdgpu_amdkfd_restore_userptr_worker()
2805 mutex_unlock(&process_info->lock); in amdgpu_amdkfd_restore_userptr_worker()
2810 &process_info->restore_userptr_work, in amdgpu_amdkfd_restore_userptr_worker()
2823 /* protected by process_info->lock */); in replace_eviction_fence()
2836 /** amdgpu_amdkfd_gpuvm_restore_process_bos - Restore all BOs for the given
2842 * should be called when the Process is still valid. BO restore involves -
2858 struct kgd_mem *mem; in amdgpu_amdkfd_gpuvm_restore_process_bos() local
2868 mutex_lock(&process_info->lock); in amdgpu_amdkfd_gpuvm_restore_process_bos()
2872 list_for_each_entry(peer_vm, &process_info->vm_list_head, in amdgpu_amdkfd_gpuvm_restore_process_bos()
2883 list_for_each_entry(mem, &process_info->kfd_bo_list, in amdgpu_amdkfd_gpuvm_restore_process_bos()
2887 gobj = &mem->bo->tbo.base; in amdgpu_amdkfd_gpuvm_restore_process_bos()
2903 list_for_each_entry(mem, &process_info->kfd_bo_list, in amdgpu_amdkfd_gpuvm_restore_process_bos()
2906 struct amdgpu_bo *bo = mem->bo; in amdgpu_amdkfd_gpuvm_restore_process_bos()
2907 uint32_t domain = mem->domain; in amdgpu_amdkfd_gpuvm_restore_process_bos()
2925 dma_resv_for_each_fence(&cursor, bo->tbo.base.resv, in amdgpu_amdkfd_gpuvm_restore_process_bos()
2933 list_for_each_entry(attachment, &mem->attachments, list) { in amdgpu_amdkfd_gpuvm_restore_process_bos()
2934 if (!attachment->is_mapped) in amdgpu_amdkfd_gpuvm_restore_process_bos()
2937 if (attachment->bo_va->base.bo->tbo.pin_count) in amdgpu_amdkfd_gpuvm_restore_process_bos()
2940 kfd_mem_dmaunmap_attachment(mem, attachment); in amdgpu_amdkfd_gpuvm_restore_process_bos()
2941 ret = update_gpuvm_pte(mem, attachment, &sync_obj); in amdgpu_amdkfd_gpuvm_restore_process_bos()
2953 list_for_each_entry(peer_vm, &process_info->vm_list_head, in amdgpu_amdkfd_gpuvm_restore_process_bos()
2956 peer_vm->root.bo->tbo.bdev); in amdgpu_amdkfd_gpuvm_restore_process_bos()
2994 if (dma_fence_is_signaled(&process_info->eviction_fence->base)) { in amdgpu_amdkfd_gpuvm_restore_process_bos()
2997 process_info->eviction_fence->base.context, in amdgpu_amdkfd_gpuvm_restore_process_bos()
2998 process_info->eviction_fence->mm, in amdgpu_amdkfd_gpuvm_restore_process_bos()
3003 ret = -ENOMEM; in amdgpu_amdkfd_gpuvm_restore_process_bos()
3006 dma_fence_put(&process_info->eviction_fence->base); in amdgpu_amdkfd_gpuvm_restore_process_bos()
3007 process_info->eviction_fence = new_fence; in amdgpu_amdkfd_gpuvm_restore_process_bos()
3008 replace_eviction_fence(ef, dma_fence_get(&new_fence->base)); in amdgpu_amdkfd_gpuvm_restore_process_bos()
3010 WARN_ONCE(*ef != &process_info->eviction_fence->base, in amdgpu_amdkfd_gpuvm_restore_process_bos()
3015 list_for_each_entry(mem, &process_info->kfd_bo_list, validate_list) { in amdgpu_amdkfd_gpuvm_restore_process_bos()
3016 if (mem->bo->tbo.pin_count) in amdgpu_amdkfd_gpuvm_restore_process_bos()
3019 dma_resv_add_fence(mem->bo->tbo.base.resv, in amdgpu_amdkfd_gpuvm_restore_process_bos()
3020 &process_info->eviction_fence->base, in amdgpu_amdkfd_gpuvm_restore_process_bos()
3024 list_for_each_entry(peer_vm, &process_info->vm_list_head, in amdgpu_amdkfd_gpuvm_restore_process_bos()
3026 struct amdgpu_bo *bo = peer_vm->root.bo; in amdgpu_amdkfd_gpuvm_restore_process_bos()
3028 dma_resv_add_fence(bo->tbo.base.resv, in amdgpu_amdkfd_gpuvm_restore_process_bos()
3029 &process_info->eviction_fence->base, in amdgpu_amdkfd_gpuvm_restore_process_bos()
3037 mutex_unlock(&process_info->lock); in amdgpu_amdkfd_gpuvm_restore_process_bos()
3041 int amdgpu_amdkfd_add_gws_to_process(void *info, void *gws, struct kgd_mem **mem) in amdgpu_amdkfd_add_gws_to_process() argument
3048 return -EINVAL; in amdgpu_amdkfd_add_gws_to_process()
3050 *mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL); in amdgpu_amdkfd_add_gws_to_process()
3051 if (!*mem) in amdgpu_amdkfd_add_gws_to_process()
3052 return -ENOMEM; in amdgpu_amdkfd_add_gws_to_process()
3054 mutex_init(&(*mem)->lock); in amdgpu_amdkfd_add_gws_to_process()
3055 INIT_LIST_HEAD(&(*mem)->attachments); in amdgpu_amdkfd_add_gws_to_process()
3056 (*mem)->bo = amdgpu_bo_ref(gws_bo); in amdgpu_amdkfd_add_gws_to_process()
3057 (*mem)->domain = AMDGPU_GEM_DOMAIN_GWS; in amdgpu_amdkfd_add_gws_to_process()
3058 (*mem)->process_info = process_info; in amdgpu_amdkfd_add_gws_to_process()
3059 add_kgd_mem_to_kfd_bo_list(*mem, process_info, false); in amdgpu_amdkfd_add_gws_to_process()
3060 amdgpu_sync_create(&(*mem)->sync); in amdgpu_amdkfd_add_gws_to_process()
3064 mutex_lock(&(*mem)->process_info->lock); in amdgpu_amdkfd_add_gws_to_process()
3080 ret = dma_resv_reserve_fences(gws_bo->tbo.base.resv, 1); in amdgpu_amdkfd_add_gws_to_process()
3083 dma_resv_add_fence(gws_bo->tbo.base.resv, in amdgpu_amdkfd_add_gws_to_process()
3084 &process_info->eviction_fence->base, in amdgpu_amdkfd_add_gws_to_process()
3087 mutex_unlock(&(*mem)->process_info->lock); in amdgpu_amdkfd_add_gws_to_process()
3095 mutex_unlock(&(*mem)->process_info->lock); in amdgpu_amdkfd_add_gws_to_process()
3096 amdgpu_sync_free(&(*mem)->sync); in amdgpu_amdkfd_add_gws_to_process()
3097 remove_kgd_mem_from_kfd_bo_list(*mem, process_info); in amdgpu_amdkfd_add_gws_to_process()
3099 mutex_destroy(&(*mem)->lock); in amdgpu_amdkfd_add_gws_to_process()
3100 kfree(*mem); in amdgpu_amdkfd_add_gws_to_process()
3101 *mem = NULL; in amdgpu_amdkfd_add_gws_to_process()
3105 int amdgpu_amdkfd_remove_gws_from_process(void *info, void *mem) in amdgpu_amdkfd_remove_gws_from_process() argument
3109 struct kgd_mem *kgd_mem = (struct kgd_mem *)mem; in amdgpu_amdkfd_remove_gws_from_process()
3110 struct amdgpu_bo *gws_bo = kgd_mem->bo; in amdgpu_amdkfd_remove_gws_from_process()
3124 process_info->eviction_fence); in amdgpu_amdkfd_remove_gws_from_process()
3126 amdgpu_sync_free(&kgd_mem->sync); in amdgpu_amdkfd_remove_gws_from_process()
3128 mutex_destroy(&kgd_mem->lock); in amdgpu_amdkfd_remove_gws_from_process()
3129 kfree(mem); in amdgpu_amdkfd_remove_gws_from_process()
3133 /* Returns GPU-specific tiling mode information */
3137 config->gb_addr_config = adev->gfx.config.gb_addr_config; in amdgpu_amdkfd_get_tile_config()
3138 config->tile_config_ptr = adev->gfx.config.tile_mode_array; in amdgpu_amdkfd_get_tile_config()
3139 config->num_tile_configs = in amdgpu_amdkfd_get_tile_config()
3140 ARRAY_SIZE(adev->gfx.config.tile_mode_array); in amdgpu_amdkfd_get_tile_config()
3141 config->macro_tile_config_ptr = in amdgpu_amdkfd_get_tile_config()
3142 adev->gfx.config.macrotile_mode_array; in amdgpu_amdkfd_get_tile_config()
3143 config->num_macro_tile_configs = in amdgpu_amdkfd_get_tile_config()
3144 ARRAY_SIZE(adev->gfx.config.macrotile_mode_array); in amdgpu_amdkfd_get_tile_config()
3147 config->num_banks = adev->gfx.config.num_banks; in amdgpu_amdkfd_get_tile_config()
3148 config->num_ranks = adev->gfx.config.num_ranks; in amdgpu_amdkfd_get_tile_config()
3153 bool amdgpu_amdkfd_bo_mapped_to_dev(struct amdgpu_device *adev, struct kgd_mem *mem) in amdgpu_amdkfd_bo_mapped_to_dev() argument
3157 list_for_each_entry(entry, &mem->attachments, list) { in amdgpu_amdkfd_bo_mapped_to_dev()
3158 if (entry->is_mapped && entry->adev == adev) in amdgpu_amdkfd_bo_mapped_to_dev()
3170 seq_printf(m, "System mem used %lldM out of %lluM\n", in kfd_debugfs_kfd_mem_limits()
3173 seq_printf(m, "TTM mem used %lldM out of %lluM\n", in kfd_debugfs_kfd_mem_limits()