diff options
Diffstat (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c')
| -rw-r--r-- | drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 1231 |
1 files changed, 766 insertions, 465 deletions
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c index d6320c836251..b1c24c8fa686 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c @@ -27,6 +27,8 @@ #include <linux/sched/task.h> #include <drm/ttm/ttm_tt.h> +#include <drm/drm_exec.h> + #include "amdgpu_object.h" #include "amdgpu_gem.h" #include "amdgpu_vm.h" @@ -35,12 +37,14 @@ #include "amdgpu_dma_buf.h" #include <uapi/linux/kfd_ioctl.h> #include "amdgpu_xgmi.h" +#include "kfd_priv.h" #include "kfd_smi_events.h" /* Userptr restore delay, just long enough to allow consecutive VM * changes to accumulate */ #define AMDGPU_USERPTR_RESTORE_DELAY_MS 1 +#define AMDGPU_RESERVE_MEM_LIMIT (3UL << 29) /* * Align VRAM availability to 2MB to avoid fragmentation caused by 4K allocations in the tail 2MB @@ -82,6 +86,25 @@ static bool kfd_mem_is_attached(struct amdgpu_vm *avm, return false; } +/** + * reuse_dmamap() - Check whether adev can share the original + * userptr BO + * + * If both adev and bo_adev are in direct mapping or + * in the same iommu group, they can share the original BO. + * + * @adev: Device to which can or cannot share the original BO + * @bo_adev: Device to which allocated BO belongs to + * + * Return: returns true if adev can share original userptr BO, + * false otherwise. + */ +static bool reuse_dmamap(struct amdgpu_device *adev, struct amdgpu_device *bo_adev) +{ + return (adev->ram_is_direct_mapped && bo_adev->ram_is_direct_mapped) || + (adev->dev->iommu_group == bo_adev->dev->iommu_group); +} + /* Set memory usage limits. Current, limits are * System (TTM + userptr) memory - 15/16th System RAM * TTM memory - 3/8th System RAM @@ -91,13 +114,21 @@ void amdgpu_amdkfd_gpuvm_init_mem_limits(void) struct sysinfo si; uint64_t mem; + if (kfd_mem_limit.max_system_mem_limit) + return; + si_meminfo(&si); - mem = si.freeram - si.freehigh; + mem = si.totalram - si.totalhigh; mem *= si.mem_unit; spin_lock_init(&kfd_mem_limit.mem_limit_lock); - kfd_mem_limit.max_system_mem_limit = mem - (mem >> 4); - kfd_mem_limit.max_ttm_mem_limit = (mem >> 1) - (mem >> 3); + kfd_mem_limit.max_system_mem_limit = mem - (mem >> 6); + if (kfd_mem_limit.max_system_mem_limit < 2 * AMDGPU_RESERVE_MEM_LIMIT) + kfd_mem_limit.max_system_mem_limit >>= 1; + else + kfd_mem_limit.max_system_mem_limit -= AMDGPU_RESERVE_MEM_LIMIT; + + kfd_mem_limit.max_ttm_mem_limit = ttm_tt_pages_limit() << PAGE_SHIFT; pr_debug("Kernel memory limit %lluM, TTM limit %lluM\n", (kfd_mem_limit.max_system_mem_limit >> 20), (kfd_mem_limit.max_ttm_mem_limit >> 20)); @@ -129,16 +160,22 @@ void amdgpu_amdkfd_reserve_system_mem(uint64_t size) * @size: Size of buffer, in bytes, encapsulated by B0. This should be * equivalent to amdgpu_bo_size(BO) * @alloc_flag: Flag used in allocating a BO as noted above + * @xcp_id: xcp_id is used to get xcp from xcp manager, one xcp is + * managed as one compute node in driver for app * - * Return: returns -ENOMEM in case of error, ZERO otherwise + * Return: + * returns -ENOMEM in case of error, ZERO otherwise */ int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev, - uint64_t size, u32 alloc_flag) + uint64_t size, u32 alloc_flag, int8_t xcp_id) { uint64_t reserved_for_pt = ESTIMATE_PT_SIZE(amdgpu_amdkfd_total_mem_size); + struct amdgpu_ras *con = amdgpu_ras_get_context(adev); + uint64_t reserved_for_ras = (con ? con->reserved_pages_in_bytes : 0); size_t system_mem_needed, ttm_mem_needed, vram_needed; int ret = 0; + uint64_t vram_size = 0; system_mem_needed = 0; ttm_mem_needed = 0; @@ -153,6 +190,17 @@ int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev, * 2M BO chunk. */ vram_needed = size; + /* + * For GFX 9.4.3, get the VRAM size from XCP structs + */ + if (WARN_ONCE(xcp_id < 0, "invalid XCP ID %d", xcp_id)) + return -EINVAL; + + vram_size = KFD_XCP_MEMORY_SIZE(adev, xcp_id); + if (adev->apu_prefer_gtt) { + system_mem_needed = size; + ttm_mem_needed = size; + } } else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) { system_mem_needed = size; } else if (!(alloc_flag & @@ -165,27 +213,46 @@ int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev, spin_lock(&kfd_mem_limit.mem_limit_lock); if (kfd_mem_limit.system_mem_used + system_mem_needed > - kfd_mem_limit.max_system_mem_limit) + kfd_mem_limit.max_system_mem_limit) { pr_debug("Set no_system_mem_limit=1 if using shared memory\n"); + if (!no_system_mem_limit) { + ret = -ENOMEM; + goto release; + } + } - if ((kfd_mem_limit.system_mem_used + system_mem_needed > - kfd_mem_limit.max_system_mem_limit && !no_system_mem_limit) || - (kfd_mem_limit.ttm_mem_used + ttm_mem_needed > - kfd_mem_limit.max_ttm_mem_limit) || - (adev && adev->kfd.vram_used + vram_needed > - adev->gmc.real_vram_size - reserved_for_pt)) { + if (kfd_mem_limit.ttm_mem_used + ttm_mem_needed > + kfd_mem_limit.max_ttm_mem_limit) { ret = -ENOMEM; goto release; } + /*if is_app_apu is false and apu_prefer_gtt is true, it is an APU with + * carve out < gtt. In that case, VRAM allocation will go to gtt domain, skip + * VRAM check since ttm_mem_limit check already cover this allocation + */ + + if (adev && xcp_id >= 0 && (!adev->apu_prefer_gtt || adev->gmc.is_app_apu)) { + uint64_t vram_available = + vram_size - reserved_for_pt - reserved_for_ras - + atomic64_read(&adev->vram_pin_size); + if (adev->kfd.vram_used[xcp_id] + vram_needed > vram_available) { + ret = -ENOMEM; + goto release; + } + } + /* Update memory accounting by decreasing available system * memory, TTM memory and GPU memory as computed above */ WARN_ONCE(vram_needed && !adev, "adev reference can't be null when vram is used"); - if (adev) { - adev->kfd.vram_used += vram_needed; - adev->kfd.vram_used_aligned += ALIGN(vram_needed, VRAM_AVAILABLITY_ALIGN); + if (adev && xcp_id >= 0) { + adev->kfd.vram_used[xcp_id] += vram_needed; + adev->kfd.vram_used_aligned[xcp_id] += + adev->apu_prefer_gtt ? + vram_needed : + ALIGN(vram_needed, VRAM_AVAILABLITY_ALIGN); } kfd_mem_limit.system_mem_used += system_mem_needed; kfd_mem_limit.ttm_mem_used += ttm_mem_needed; @@ -196,7 +263,7 @@ release: } void amdgpu_amdkfd_unreserve_mem_limit(struct amdgpu_device *adev, - uint64_t size, u32 alloc_flag) + uint64_t size, u32 alloc_flag, int8_t xcp_id) { spin_lock(&kfd_mem_limit.mem_limit_lock); @@ -206,9 +273,19 @@ void amdgpu_amdkfd_unreserve_mem_limit(struct amdgpu_device *adev, } else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) { WARN_ONCE(!adev, "adev reference can't be null when alloc mem flags vram is set"); + if (WARN_ONCE(xcp_id < 0, "invalid XCP ID %d", xcp_id)) + goto release; + if (adev) { - adev->kfd.vram_used -= size; - adev->kfd.vram_used_aligned -= ALIGN(size, VRAM_AVAILABLITY_ALIGN); + adev->kfd.vram_used[xcp_id] -= size; + if (adev->apu_prefer_gtt) { + adev->kfd.vram_used_aligned[xcp_id] -= size; + kfd_mem_limit.system_mem_used -= size; + kfd_mem_limit.ttm_mem_used -= size; + } else { + adev->kfd.vram_used_aligned[xcp_id] -= + ALIGN(size, VRAM_AVAILABLITY_ALIGN); + } } } else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) { kfd_mem_limit.system_mem_used -= size; @@ -218,8 +295,8 @@ void amdgpu_amdkfd_unreserve_mem_limit(struct amdgpu_device *adev, pr_err("%s: Invalid BO type %#x\n", __func__, alloc_flag); goto release; } - WARN_ONCE(adev && adev->kfd.vram_used < 0, - "KFD VRAM memory accounting unbalanced"); + WARN_ONCE(adev && xcp_id >= 0 && adev->kfd.vram_used[xcp_id] < 0, + "KFD VRAM memory accounting unbalanced for xcp: %d", xcp_id); WARN_ONCE(kfd_mem_limit.ttm_mem_used < 0, "KFD TTM memory accounting unbalanced"); WARN_ONCE(kfd_mem_limit.system_mem_used < 0, @@ -235,14 +312,16 @@ void amdgpu_amdkfd_release_notify(struct amdgpu_bo *bo) u32 alloc_flags = bo->kfd_bo->alloc_flags; u64 size = amdgpu_bo_size(bo); - amdgpu_amdkfd_unreserve_mem_limit(adev, size, alloc_flags); + amdgpu_amdkfd_unreserve_mem_limit(adev, size, alloc_flags, + bo->xcp_id); kfree(bo->kfd_bo); } /** - * @create_dmamap_sg_bo: Creates a amdgpu_bo object to reflect information + * create_dmamap_sg_bo() - Creates a amdgpu_bo object to reflect information * about USERPTR or DOOREBELL or MMIO BO. + * * @adev: Device for which dmamap BO is being created * @mem: BO of peer device that is being DMA mapped. Provides parameters * in building the dmamap BO @@ -253,16 +332,20 @@ create_dmamap_sg_bo(struct amdgpu_device *adev, struct kgd_mem *mem, struct amdgpu_bo **bo_out) { struct drm_gem_object *gem_obj; - int ret, align; + int ret; + uint64_t flags = 0; ret = amdgpu_bo_reserve(mem->bo, false); if (ret) return ret; - align = 1; - ret = amdgpu_gem_object_create(adev, mem->bo->tbo.base.size, align, - AMDGPU_GEM_DOMAIN_CPU, AMDGPU_GEM_CREATE_PREEMPTIBLE, - ttm_bo_type_sg, mem->bo->tbo.base.resv, &gem_obj); + if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) + flags |= mem->bo->flags & (AMDGPU_GEM_CREATE_COHERENT | + AMDGPU_GEM_CREATE_UNCACHED); + + ret = amdgpu_gem_object_create(adev, mem->bo->tbo.base.size, 1, + AMDGPU_GEM_DOMAIN_CPU, AMDGPU_GEM_CREATE_PREEMPTIBLE | flags, + ttm_bo_type_sg, mem->bo->tbo.base.resv, &gem_obj, 0); amdgpu_bo_unreserve(mem->bo); @@ -303,40 +386,32 @@ static int amdgpu_amdkfd_remove_eviction_fence(struct amdgpu_bo *bo, return 0; } -int amdgpu_amdkfd_remove_fence_on_pt_pd_bos(struct amdgpu_bo *bo) +/** + * amdgpu_amdkfd_remove_all_eviction_fences - Remove all eviction fences + * @bo: the BO where to remove the evictions fences from. + * + * This functions should only be used on release when all references to the BO + * are already dropped. We remove the eviction fence from the private copy of + * the dma_resv object here since that is what is used during release to + * determine of the BO is idle or not. + */ +void amdgpu_amdkfd_remove_all_eviction_fences(struct amdgpu_bo *bo) { - struct amdgpu_bo *root = bo; - struct amdgpu_vm_bo_base *vm_bo; - struct amdgpu_vm *vm; - struct amdkfd_process_info *info; - struct amdgpu_amdkfd_fence *ef; - int ret; - - /* we can always get vm_bo from root PD bo.*/ - while (root->parent) - root = root->parent; - - vm_bo = root->vm_bo; - if (!vm_bo) - return 0; - - vm = vm_bo->vm; - if (!vm) - return 0; + struct dma_resv *resv = &bo->tbo.base._resv; + struct dma_fence *fence, *stub; + struct dma_resv_iter cursor; - info = vm->process_info; - if (!info || !info->eviction_fence) - return 0; - - ef = container_of(dma_fence_get(&info->eviction_fence->base), - struct amdgpu_amdkfd_fence, base); + dma_resv_assert_held(resv); - BUG_ON(!dma_resv_trylock(bo->tbo.base.resv)); - ret = amdgpu_amdkfd_remove_eviction_fence(bo, ef); - dma_resv_unlock(bo->tbo.base.resv); + stub = dma_fence_get_stub(); + dma_resv_for_each_fence(&cursor, resv, DMA_RESV_USAGE_BOOKKEEP, fence) { + if (!to_amdgpu_amdkfd_fence(fence)) + continue; - dma_fence_put(&ef->base); - return ret; + dma_resv_replace_fences(resv, fence->context, stub, + DMA_RESV_USAGE_BOOKKEEP); + } + dma_fence_put(stub); } static int amdgpu_amdkfd_bo_validate(struct amdgpu_bo *bo, uint32_t domain, @@ -349,6 +424,10 @@ static int amdgpu_amdkfd_bo_validate(struct amdgpu_bo *bo, uint32_t domain, "Called with userptr BO")) return -EINVAL; + /* bo has been pinned, not need validate it */ + if (bo->tbo.pin_count) + return 0; + amdgpu_bo_placement_from_domain(bo, domain); ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); @@ -361,6 +440,32 @@ validate_fail: return ret; } +int amdgpu_amdkfd_bo_validate_and_fence(struct amdgpu_bo *bo, + uint32_t domain, + struct dma_fence *fence) +{ + int ret = amdgpu_bo_reserve(bo, false); + + if (ret) + return ret; + + ret = amdgpu_amdkfd_bo_validate(bo, domain, true); + if (ret) + goto unreserve_out; + + ret = dma_resv_reserve_fences(bo->tbo.base.resv, 1); + if (ret) + goto unreserve_out; + + dma_resv_add_fence(bo->tbo.base.resv, fence, + DMA_RESV_USAGE_BOOKKEEP); + +unreserve_out: + amdgpu_bo_unreserve(bo); + + return ret; +} + static int amdgpu_amdkfd_validate_vm_bo(void *_unused, struct amdgpu_bo *bo) { return amdgpu_amdkfd_bo_validate(bo, bo->allowed_domains, false); @@ -373,13 +478,15 @@ static int amdgpu_amdkfd_validate_vm_bo(void *_unused, struct amdgpu_bo *bo) * again. Page directories are only updated after updating page * tables. */ -static int vm_validate_pt_pd_bos(struct amdgpu_vm *vm) +static int vm_validate_pt_pd_bos(struct amdgpu_vm *vm, + struct ww_acquire_ctx *ticket) { struct amdgpu_bo *pd = vm->root.bo; struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev); int ret; - ret = amdgpu_vm_validate_pt_bos(adev, vm, amdgpu_amdkfd_validate_vm_bo, NULL); + ret = amdgpu_vm_validate(adev, vm, ticket, + amdgpu_amdkfd_validate_vm_bo, NULL); if (ret) { pr_err("failed to validate PT BOs\n"); return ret; @@ -400,10 +507,11 @@ static int vm_update_pds(struct amdgpu_vm *vm, struct amdgpu_sync *sync) if (ret) return ret; - return amdgpu_sync_fence(sync, vm->last_update); + return amdgpu_sync_fence(sync, vm->last_update, GFP_KERNEL); } -static uint64_t get_pte_flags(struct amdgpu_device *adev, struct kgd_mem *mem) +static uint64_t get_pte_flags(struct amdgpu_device *adev, struct amdgpu_vm *vm, + struct kgd_mem *mem) { uint32_t mapping_flags = AMDGPU_VM_PAGE_READABLE | AMDGPU_VM_MTYPE_DEFAULT; @@ -413,7 +521,7 @@ static uint64_t get_pte_flags(struct amdgpu_device *adev, struct kgd_mem *mem) if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE) mapping_flags |= AMDGPU_VM_PAGE_EXECUTABLE; - return amdgpu_gem_va_map_flags(adev, mapping_flags); + return mapping_flags; } /** @@ -481,9 +589,6 @@ kfd_mem_dmamap_userptr(struct kgd_mem *mem, if (unlikely(ret)) goto release_sg; - drm_prime_sg_to_dma_addr_array(ttm->sg, ttm->dma_address, - ttm->num_pages); - amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT); ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); if (ret) @@ -628,7 +733,7 @@ kfd_mem_dmaunmap_userptr(struct kgd_mem *mem, return; amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU); - ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); + (void)ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); dma_unmap_sgtable(adev->dev, ttm->sg, direction, 0); sg_free_table(ttm->sg); @@ -639,11 +744,10 @@ kfd_mem_dmaunmap_userptr(struct kgd_mem *mem, static void kfd_mem_dmaunmap_dmabuf(struct kfd_mem_attachment *attachment) { - struct ttm_operation_ctx ctx = {.interruptible = true}; - struct amdgpu_bo *bo = attachment->bo_va->base.bo; - - amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU); - ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); + /* This is a no-op. We don't want to trigger eviction fences when + * unmapping DMABufs. Therefore the invalidation (moving to system + * domain) is done in kfd_mem_dmamap_dmabuf. + */ } /** @@ -673,12 +777,12 @@ kfd_mem_dmaunmap_sg_bo(struct kgd_mem *mem, enum dma_data_direction dir; if (unlikely(!ttm->sg)) { - pr_err("SG Table of BO is UNEXPECTEDLY NULL"); + pr_debug("SG Table of BO is NULL"); return; } amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU); - ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); + (void)ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); dir = mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ? DMA_BIDIRECTIONAL : DMA_TO_DEVICE; @@ -711,6 +815,25 @@ kfd_mem_dmaunmap_attachment(struct kgd_mem *mem, } } +static int kfd_mem_export_dmabuf(struct kgd_mem *mem) +{ + if (!mem->dmabuf) { + struct amdgpu_device *bo_adev; + struct dma_buf *dmabuf; + + bo_adev = amdgpu_ttm_adev(mem->bo->tbo.bdev); + dmabuf = drm_gem_prime_handle_to_dmabuf(&bo_adev->ddev, bo_adev->kfd.client.file, + mem->gem_handle, + mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ? + DRM_RDWR : 0); + if (IS_ERR(dmabuf)) + return PTR_ERR(dmabuf); + mem->dmabuf = dmabuf; + } + + return 0; +} + static int kfd_mem_attach_dmabuf(struct amdgpu_device *adev, struct kgd_mem *mem, struct amdgpu_bo **bo) @@ -718,16 +841,9 @@ kfd_mem_attach_dmabuf(struct amdgpu_device *adev, struct kgd_mem *mem, struct drm_gem_object *gobj; int ret; - if (!mem->dmabuf) { - mem->dmabuf = amdgpu_gem_prime_export(&mem->bo->tbo.base, - mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ? - DRM_RDWR : 0); - if (IS_ERR(mem->dmabuf)) { - ret = PTR_ERR(mem->dmabuf); - mem->dmabuf = NULL; - return ret; - } - } + ret = kfd_mem_export_dmabuf(mem); + if (ret) + return ret; gobj = amdgpu_gem_prime_import(adev_to_drm(adev), mem->dmabuf); if (IS_ERR(gobj)) @@ -760,6 +876,7 @@ static int kfd_mem_attach(struct amdgpu_device *adev, struct kgd_mem *mem, uint64_t va = mem->va; struct kfd_mem_attachment *attachment[2] = {NULL, NULL}; struct amdgpu_bo *bo[2] = {NULL, NULL}; + struct amdgpu_bo_va *bo_va; bool same_hive = false; int i, ret; @@ -776,7 +893,7 @@ static int kfd_mem_attach(struct amdgpu_device *adev, struct kgd_mem *mem, * if peer device has large BAR. In contrast, access over xGMI is * allowed for both small and large BAR configurations of peer device */ - if ((adev != bo_adev) && + if ((adev != bo_adev && !adev->apu_prefer_gtt) && ((mem->domain == AMDGPU_GEM_DOMAIN_VRAM) || (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) || (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP))) { @@ -797,11 +914,12 @@ static int kfd_mem_attach(struct amdgpu_device *adev, struct kgd_mem *mem, va + bo_size, vm); if ((adev == bo_adev && !(mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) || - (amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm) && adev->ram_is_direct_mapped) || + (amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm) && reuse_dmamap(adev, bo_adev)) || + (mem->domain == AMDGPU_GEM_DOMAIN_GTT && reuse_dmamap(adev, bo_adev)) || same_hive) { /* Mappings on the local GPU, or VRAM mappings in the - * local hive, or userptr mapping IOMMU direct map mode - * share the original BO + * local hive, or userptr, or GTT mapping can reuse dma map + * address space share the original BO */ attachment[i]->type = KFD_MEM_ATT_SHARED; bo[i] = mem->bo; @@ -846,7 +964,12 @@ static int kfd_mem_attach(struct amdgpu_device *adev, struct kgd_mem *mem, pr_debug("Unable to reserve BO during memory attach"); goto unwind; } - attachment[i]->bo_va = amdgpu_vm_bo_add(adev, vm, bo[i]); + bo_va = amdgpu_vm_bo_find(vm, bo[i]); + if (!bo_va) + bo_va = amdgpu_vm_bo_add(adev, vm, bo[i]); + else + ++bo_va->ref_count; + attachment[i]->bo_va = bo_va; amdgpu_bo_unreserve(bo[i]); if (unlikely(!attachment[i]->bo_va)) { ret = -ENOMEM; @@ -855,7 +978,7 @@ static int kfd_mem_attach(struct amdgpu_device *adev, struct kgd_mem *mem, goto unwind; } attachment[i]->va = va; - attachment[i]->pte_flags = get_pte_flags(adev, mem); + attachment[i]->pte_flags = get_pte_flags(adev, vm, mem); attachment[i]->adev = adev; list_add(&attachment[i]->list, &mem->attachments); @@ -869,8 +992,9 @@ unwind: if (!attachment[i]) continue; if (attachment[i]->bo_va) { - amdgpu_bo_reserve(bo[i], true); - amdgpu_vm_bo_del(adev, attachment[i]->bo_va); + (void)amdgpu_bo_reserve(bo[i], true); + if (--attachment[i]->bo_va->ref_count == 0) + amdgpu_vm_bo_del(adev, attachment[i]->bo_va); amdgpu_bo_unreserve(bo[i]); list_del(&attachment[i]->list); } @@ -887,7 +1011,8 @@ static void kfd_mem_detach(struct kfd_mem_attachment *attachment) pr_debug("\t remove VA 0x%llx in entry %p\n", attachment->va, attachment); - amdgpu_vm_bo_del(attachment->adev, attachment->bo_va); + if (--attachment->bo_va->ref_count == 0) + amdgpu_vm_bo_del(attachment->adev, attachment->bo_va); drm_gem_object_put(&bo->tbo.base); list_del(&attachment->list); kfree(attachment); @@ -897,28 +1022,20 @@ static void add_kgd_mem_to_kfd_bo_list(struct kgd_mem *mem, struct amdkfd_process_info *process_info, bool userptr) { - struct ttm_validate_buffer *entry = &mem->validate_list; - struct amdgpu_bo *bo = mem->bo; - - INIT_LIST_HEAD(&entry->head); - entry->num_shared = 1; - entry->bo = &bo->tbo; mutex_lock(&process_info->lock); if (userptr) - list_add_tail(&entry->head, &process_info->userptr_valid_list); + list_add_tail(&mem->validate_list, + &process_info->userptr_valid_list); else - list_add_tail(&entry->head, &process_info->kfd_bo_list); + list_add_tail(&mem->validate_list, &process_info->kfd_bo_list); mutex_unlock(&process_info->lock); } static void remove_kgd_mem_from_kfd_bo_list(struct kgd_mem *mem, struct amdkfd_process_info *process_info) { - struct ttm_validate_buffer *bo_list_entry; - - bo_list_entry = &mem->validate_list; mutex_lock(&process_info->lock); - list_del(&bo_list_entry->head); + list_del(&mem->validate_list); mutex_unlock(&process_info->lock); } @@ -940,7 +1057,7 @@ static int init_user_pages(struct kgd_mem *mem, uint64_t user_addr, struct amdkfd_process_info *process_info = mem->process_info; struct amdgpu_bo *bo = mem->bo; struct ttm_operation_ctx ctx = { true, false }; - struct hmm_range *range; + struct amdgpu_hmm_range *range; int ret = 0; mutex_lock(&process_info->lock); @@ -972,9 +1089,19 @@ static int init_user_pages(struct kgd_mem *mem, uint64_t user_addr, return 0; } - ret = amdgpu_ttm_tt_get_user_pages(bo, bo->tbo.ttm->pages, &range); + range = amdgpu_hmm_range_alloc(NULL); + if (unlikely(!range)) { + ret = -ENOMEM; + goto unregister_out; + } + + ret = amdgpu_ttm_tt_get_user_pages(bo, range); if (ret) { - pr_err("%s: Failed to get user pages: %d\n", __func__, ret); + amdgpu_hmm_range_free(range); + if (ret == -EAGAIN) + pr_debug("Failed to get user pages, try again\n"); + else + pr_err("%s: Failed to get user pages: %d\n", __func__, ret); goto unregister_out; } @@ -983,6 +1110,9 @@ static int init_user_pages(struct kgd_mem *mem, uint64_t user_addr, pr_err("%s: Failed to reserve BO\n", __func__); goto release_out; } + + amdgpu_ttm_tt_set_user_pages(bo->tbo.ttm, range); + amdgpu_bo_placement_from_domain(bo, mem->domain); ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); if (ret) @@ -990,7 +1120,7 @@ static int init_user_pages(struct kgd_mem *mem, uint64_t user_addr, amdgpu_bo_unreserve(bo); release_out: - amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm, range); + amdgpu_hmm_range_free(range); unregister_out: if (ret) amdgpu_hmm_unregister(bo); @@ -1005,13 +1135,12 @@ out: * object can track VM updates. */ struct bo_vm_reservation_context { - struct amdgpu_bo_list_entry kfd_bo; /* BO list entry for the KFD BO */ - unsigned int n_vms; /* Number of VMs reserved */ - struct amdgpu_bo_list_entry *vm_pd; /* Array of VM BO list entries */ - struct ww_acquire_ctx ticket; /* Reservation ticket */ - struct list_head list, duplicates; /* BO lists */ - struct amdgpu_sync *sync; /* Pointer to sync object */ - bool reserved; /* Whether BOs are reserved */ + /* DRM execution context for the reservation */ + struct drm_exec exec; + /* Number of VMs reserved */ + unsigned int n_vms; + /* Pointer to sync object */ + struct amdgpu_sync *sync; }; enum bo_vm_match { @@ -1035,35 +1164,26 @@ static int reserve_bo_and_vm(struct kgd_mem *mem, WARN_ON(!vm); - ctx->reserved = false; ctx->n_vms = 1; ctx->sync = &mem->sync; - - INIT_LIST_HEAD(&ctx->list); - INIT_LIST_HEAD(&ctx->duplicates); - - ctx->vm_pd = kcalloc(ctx->n_vms, sizeof(*ctx->vm_pd), GFP_KERNEL); - if (!ctx->vm_pd) - return -ENOMEM; - - ctx->kfd_bo.priority = 0; - ctx->kfd_bo.tv.bo = &bo->tbo; - ctx->kfd_bo.tv.num_shared = 1; - list_add(&ctx->kfd_bo.tv.head, &ctx->list); - - amdgpu_vm_get_pd_bo(vm, &ctx->list, &ctx->vm_pd[0]); - - ret = ttm_eu_reserve_buffers(&ctx->ticket, &ctx->list, - false, &ctx->duplicates); - if (ret) { - pr_err("Failed to reserve buffers in ttm.\n"); - kfree(ctx->vm_pd); - ctx->vm_pd = NULL; - return ret; + drm_exec_init(&ctx->exec, DRM_EXEC_INTERRUPTIBLE_WAIT, 0); + drm_exec_until_all_locked(&ctx->exec) { + ret = amdgpu_vm_lock_pd(vm, &ctx->exec, 2); + drm_exec_retry_on_contention(&ctx->exec); + if (unlikely(ret)) + goto error; + + ret = drm_exec_prepare_obj(&ctx->exec, &bo->tbo.base, 1); + drm_exec_retry_on_contention(&ctx->exec); + if (unlikely(ret)) + goto error; } - - ctx->reserved = true; return 0; + +error: + pr_err("Failed to reserve buffers in ttm.\n"); + drm_exec_fini(&ctx->exec); + return ret; } /** @@ -1080,63 +1200,40 @@ static int reserve_bo_and_cond_vms(struct kgd_mem *mem, struct amdgpu_vm *vm, enum bo_vm_match map_type, struct bo_vm_reservation_context *ctx) { - struct amdgpu_bo *bo = mem->bo; struct kfd_mem_attachment *entry; - unsigned int i; + struct amdgpu_bo *bo = mem->bo; int ret; - ctx->reserved = false; - ctx->n_vms = 0; - ctx->vm_pd = NULL; ctx->sync = &mem->sync; + drm_exec_init(&ctx->exec, DRM_EXEC_INTERRUPTIBLE_WAIT | + DRM_EXEC_IGNORE_DUPLICATES, 0); + drm_exec_until_all_locked(&ctx->exec) { + ctx->n_vms = 0; + list_for_each_entry(entry, &mem->attachments, list) { + if ((vm && vm != entry->bo_va->base.vm) || + (entry->is_mapped != map_type + && map_type != BO_VM_ALL)) + continue; - INIT_LIST_HEAD(&ctx->list); - INIT_LIST_HEAD(&ctx->duplicates); - - list_for_each_entry(entry, &mem->attachments, list) { - if ((vm && vm != entry->bo_va->base.vm) || - (entry->is_mapped != map_type - && map_type != BO_VM_ALL)) - continue; - - ctx->n_vms++; - } - - if (ctx->n_vms != 0) { - ctx->vm_pd = kcalloc(ctx->n_vms, sizeof(*ctx->vm_pd), - GFP_KERNEL); - if (!ctx->vm_pd) - return -ENOMEM; - } - - ctx->kfd_bo.priority = 0; - ctx->kfd_bo.tv.bo = &bo->tbo; - ctx->kfd_bo.tv.num_shared = 1; - list_add(&ctx->kfd_bo.tv.head, &ctx->list); - - i = 0; - list_for_each_entry(entry, &mem->attachments, list) { - if ((vm && vm != entry->bo_va->base.vm) || - (entry->is_mapped != map_type - && map_type != BO_VM_ALL)) - continue; - - amdgpu_vm_get_pd_bo(entry->bo_va->base.vm, &ctx->list, - &ctx->vm_pd[i]); - i++; - } + ret = amdgpu_vm_lock_pd(entry->bo_va->base.vm, + &ctx->exec, 2); + drm_exec_retry_on_contention(&ctx->exec); + if (unlikely(ret)) + goto error; + ++ctx->n_vms; + } - ret = ttm_eu_reserve_buffers(&ctx->ticket, &ctx->list, - false, &ctx->duplicates); - if (ret) { - pr_err("Failed to reserve buffers in ttm.\n"); - kfree(ctx->vm_pd); - ctx->vm_pd = NULL; - return ret; + ret = drm_exec_prepare_obj(&ctx->exec, &bo->tbo.base, 1); + drm_exec_retry_on_contention(&ctx->exec); + if (unlikely(ret)) + goto error; } - - ctx->reserved = true; return 0; + +error: + pr_err("Failed to reserve buffers in ttm.\n"); + drm_exec_fini(&ctx->exec); + return ret; } /** @@ -1157,19 +1254,12 @@ static int unreserve_bo_and_vms(struct bo_vm_reservation_context *ctx, if (wait) ret = amdgpu_sync_wait(ctx->sync, intr); - if (ctx->reserved) - ttm_eu_backoff_reservation(&ctx->ticket, &ctx->list); - kfree(ctx->vm_pd); - + drm_exec_fini(&ctx->exec); ctx->sync = NULL; - - ctx->reserved = false; - ctx->vm_pd = NULL; - return ret; } -static void unmap_bo_from_gpuvm(struct kgd_mem *mem, +static int unmap_bo_from_gpuvm(struct kgd_mem *mem, struct kfd_mem_attachment *entry, struct amdgpu_sync *sync) { @@ -1177,13 +1267,22 @@ static void unmap_bo_from_gpuvm(struct kgd_mem *mem, struct amdgpu_device *adev = entry->adev; struct amdgpu_vm *vm = bo_va->base.vm; - amdgpu_vm_bo_unmap(adev, bo_va, entry->va); + if (bo_va->queue_refcount) { + pr_debug("bo_va->queue_refcount %d\n", bo_va->queue_refcount); + return -EBUSY; + } + + (void)amdgpu_vm_bo_unmap(adev, bo_va, entry->va); - amdgpu_vm_clear_freed(adev, vm, &bo_va->last_pt_update); + /* VM entity stopped if process killed, don't clear freed pt bo */ + if (!amdgpu_vm_ready(vm)) + return 0; - amdgpu_sync_fence(sync, bo_va->last_pt_update); + (void)amdgpu_vm_clear_freed(adev, vm, &bo_va->last_pt_update); - kfd_mem_dmaunmap_attachment(mem, entry); + (void)amdgpu_sync_fence(sync, bo_va->last_pt_update, GFP_KERNEL); + + return 0; } static int update_gpuvm_pte(struct kgd_mem *mem, @@ -1205,7 +1304,7 @@ static int update_gpuvm_pte(struct kgd_mem *mem, return ret; } - return amdgpu_sync_fence(sync, bo_va->last_pt_update); + return amdgpu_sync_fence(sync, bo_va->last_pt_update, GFP_KERNEL); } static int map_bo_to_gpuvm(struct kgd_mem *mem, @@ -1238,17 +1337,19 @@ static int map_bo_to_gpuvm(struct kgd_mem *mem, update_gpuvm_pte_failed: unmap_bo_from_gpuvm(mem, entry, sync); + kfd_mem_dmaunmap_attachment(mem, entry); return ret; } -static int process_validate_vms(struct amdkfd_process_info *process_info) +static int process_validate_vms(struct amdkfd_process_info *process_info, + struct ww_acquire_ctx *ticket) { struct amdgpu_vm *peer_vm; int ret; list_for_each_entry(peer_vm, &process_info->vm_list_head, vm_list_node) { - ret = vm_validate_pt_pd_bos(peer_vm); + ret = vm_validate_pt_pd_bos(peer_vm, ticket); if (ret) return ret; } @@ -1325,7 +1426,6 @@ static int init_kfd_vm(struct amdgpu_vm *vm, void **process_info, amdgpu_amdkfd_restore_userptr_worker); *process_info = info; - *ef = dma_fence_get(&info->eviction_fence->base); } vm->process_info = *process_info; @@ -1334,7 +1434,7 @@ static int init_kfd_vm(struct amdgpu_vm *vm, void **process_info, ret = amdgpu_bo_reserve(vm->root.bo, true); if (ret) goto reserve_pd_fail; - ret = vm_validate_pt_pd_bos(vm); + ret = vm_validate_pt_pd_bos(vm, NULL); if (ret) { pr_err("validate_pt_pd_bos() failed\n"); goto validate_pd_fail; @@ -1356,6 +1456,8 @@ static int init_kfd_vm(struct amdgpu_vm *vm, void **process_info, list_add_tail(&vm->vm_list_node, &(vm->process_info->vm_list_head)); vm->process_info->n_vms++; + if (ef) + *ef = dma_fence_get(&vm->process_info->eviction_fence->base); mutex_unlock(&vm->process_info->lock); return 0; @@ -1367,10 +1469,7 @@ validate_pd_fail: reserve_pd_fail: vm->process_info = NULL; if (info) { - /* Two fence references: one in info and one in *ef */ dma_fence_put(&info->eviction_fence->base); - dma_fence_put(*ef); - *ef = NULL; *process_info = NULL; put_pid(info->pid); create_evict_fence_fail: @@ -1400,13 +1499,30 @@ static int amdgpu_amdkfd_gpuvm_pin_bo(struct amdgpu_bo *bo, u32 domain) if (unlikely(ret)) return ret; - ret = amdgpu_bo_pin_restricted(bo, domain, 0, 0); + if (bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS) { + /* + * If bo is not contiguous on VRAM, move to system memory first to ensure + * we can get contiguous VRAM space after evicting other BOs. + */ + if (!(bo->tbo.resource->placement & TTM_PL_FLAG_CONTIGUOUS)) { + struct ttm_operation_ctx ctx = { true, false }; + + amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT); + ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); + if (unlikely(ret)) { + pr_debug("validate bo 0x%p to GTT failed %d\n", &bo->tbo, ret); + goto out; + } + } + } + + ret = amdgpu_bo_pin(bo, domain); if (ret) pr_err("Error in Pinning BO to domain: %d\n", domain); amdgpu_bo_sync_wait(bo, AMDGPU_FENCE_OWNER_KFD, false); +out: amdgpu_bo_unreserve(bo); - return ret; } @@ -1430,27 +1546,6 @@ static void amdgpu_amdkfd_gpuvm_unpin_bo(struct amdgpu_bo *bo) amdgpu_bo_unreserve(bo); } -int amdgpu_amdkfd_gpuvm_set_vm_pasid(struct amdgpu_device *adev, - struct amdgpu_vm *avm, u32 pasid) - -{ - int ret; - - /* Free the original amdgpu allocated pasid, - * will be replaced with kfd allocated pasid. - */ - if (avm->pasid) { - amdgpu_pasid_free(avm->pasid); - amdgpu_vm_set_pasid(adev, avm, 0); - } - - ret = amdgpu_vm_set_pasid(adev, avm, pasid); - if (ret) - return ret; - - return 0; -} - int amdgpu_amdkfd_gpuvm_acquire_process_vm(struct amdgpu_device *adev, struct amdgpu_vm *avm, void **process_info, @@ -1508,27 +1603,6 @@ void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev, } } -void amdgpu_amdkfd_gpuvm_release_process_vm(struct amdgpu_device *adev, - void *drm_priv) -{ - struct amdgpu_vm *avm; - - if (WARN_ON(!adev || !drm_priv)) - return; - - avm = drm_priv_to_vm(drm_priv); - - pr_debug("Releasing process vm %p\n", avm); - - /* The original pasid of amdgpu vm has already been - * released during making a amdgpu vm to a compute vm - * The current pasid is managed by kfd and will be - * released on kfd process destroy. Set amdgpu pasid - * to 0 to avoid duplicate release. - */ - amdgpu_vm_release_compute(adev, avm); -} - uint64_t amdgpu_amdkfd_gpuvm_get_process_page_dir(void *drm_priv) { struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv); @@ -1564,27 +1638,57 @@ int amdgpu_amdkfd_criu_resume(void *p) goto out_unlock; } WRITE_ONCE(pinfo->block_mmu_notifications, false); - schedule_delayed_work(&pinfo->restore_userptr_work, 0); + queue_delayed_work(system_freezable_wq, + &pinfo->restore_userptr_work, 0); out_unlock: mutex_unlock(&pinfo->lock); return ret; } -size_t amdgpu_amdkfd_get_available_memory(struct amdgpu_device *adev) +size_t amdgpu_amdkfd_get_available_memory(struct amdgpu_device *adev, + uint8_t xcp_id) { uint64_t reserved_for_pt = ESTIMATE_PT_SIZE(amdgpu_amdkfd_total_mem_size); - size_t available; + struct amdgpu_ras *con = amdgpu_ras_get_context(adev); + uint64_t reserved_for_ras = (con ? con->reserved_pages_in_bytes : 0); + ssize_t available; + uint64_t vram_available, system_mem_available, ttm_mem_available; spin_lock(&kfd_mem_limit.mem_limit_lock); - available = adev->gmc.real_vram_size - - adev->kfd.vram_used_aligned - - atomic64_read(&adev->vram_pin_size) - - reserved_for_pt; + if (adev->apu_prefer_gtt && !adev->gmc.is_app_apu) + vram_available = KFD_XCP_MEMORY_SIZE(adev, xcp_id) + - adev->kfd.vram_used_aligned[xcp_id]; + else + vram_available = KFD_XCP_MEMORY_SIZE(adev, xcp_id) + - adev->kfd.vram_used_aligned[xcp_id] + - atomic64_read(&adev->vram_pin_size) + - reserved_for_pt + - reserved_for_ras; + + if (adev->apu_prefer_gtt) { + system_mem_available = no_system_mem_limit ? + kfd_mem_limit.max_system_mem_limit : + kfd_mem_limit.max_system_mem_limit - + kfd_mem_limit.system_mem_used; + + ttm_mem_available = kfd_mem_limit.max_ttm_mem_limit - + kfd_mem_limit.ttm_mem_used; + + available = min3(system_mem_available, ttm_mem_available, + vram_available); + available = ALIGN_DOWN(available, PAGE_SIZE); + } else { + available = ALIGN_DOWN(vram_available, VRAM_AVAILABLITY_ALIGN); + } + spin_unlock(&kfd_mem_limit.mem_limit_lock); - return ALIGN_DOWN(available, VRAM_AVAILABLITY_ALIGN); + if (available < 0) + available = 0; + + return available; } int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu( @@ -1593,6 +1697,7 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu( uint64_t *offset, uint32_t flags, bool criu_resume) { struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv); + struct amdgpu_fpriv *fpriv = container_of(avm, struct amdgpu_fpriv, vm); enum ttm_bo_type bo_type = ttm_bo_type_device; struct sg_table *sg = NULL; uint64_t user_addr = 0; @@ -1600,6 +1705,7 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu( struct drm_gem_object *gobj = NULL; u32 domain, alloc_domain; uint64_t aligned_size; + int8_t xcp_id = -1; u64 alloc_flags; int ret; @@ -1608,9 +1714,22 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu( */ if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) { domain = alloc_domain = AMDGPU_GEM_DOMAIN_VRAM; - alloc_flags = AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE; - alloc_flags |= (flags & KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC) ? + + if (adev->apu_prefer_gtt) { + domain = AMDGPU_GEM_DOMAIN_GTT; + alloc_domain = AMDGPU_GEM_DOMAIN_GTT; + alloc_flags = 0; + } else { + alloc_flags = AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE; + alloc_flags |= (flags & KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC) ? AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED : 0; + + /* For contiguous VRAM allocation */ + if (flags & KFD_IOC_ALLOC_MEM_FLAGS_CONTIGUOUS) + alloc_flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS; + } + xcp_id = fpriv->xcp_id == AMDGPU_XCP_NO_PARTITION ? + 0 : fpriv->xcp_id; } else if (flags & KFD_IOC_ALLOC_MEM_FLAGS_GTT) { domain = alloc_domain = AMDGPU_GEM_DOMAIN_GTT; alloc_flags = 0; @@ -1638,6 +1757,8 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu( if (flags & KFD_IOC_ALLOC_MEM_FLAGS_COHERENT) alloc_flags |= AMDGPU_GEM_CREATE_COHERENT; + if (flags & KFD_IOC_ALLOC_MEM_FLAGS_EXT_COHERENT) + alloc_flags |= AMDGPU_GEM_CREATE_EXT_COHERENT; if (flags & KFD_IOC_ALLOC_MEM_FLAGS_UNCACHED) alloc_flags |= AMDGPU_GEM_CREATE_UNCACHED; @@ -1662,17 +1783,19 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu( amdgpu_sync_create(&(*mem)->sync); - ret = amdgpu_amdkfd_reserve_mem_limit(adev, aligned_size, flags); + ret = amdgpu_amdkfd_reserve_mem_limit(adev, aligned_size, flags, + xcp_id); if (ret) { pr_debug("Insufficient memory\n"); goto err_reserve_limit; } - pr_debug("\tcreate BO VA 0x%llx size 0x%llx domain %s\n", - va, (*mem)->aql_queue ? size << 1 : size, domain_string(alloc_domain)); + pr_debug("\tcreate BO VA 0x%llx size 0x%llx domain %s xcp_id %d\n", + va, (*mem)->aql_queue ? size << 1 : size, + domain_string(alloc_domain), xcp_id); ret = amdgpu_gem_object_create(adev, aligned_size, 1, alloc_domain, alloc_flags, - bo_type, NULL, &gobj); + bo_type, NULL, &gobj, xcp_id + 1); if (ret) { pr_debug("Failed to create BO on domain %s. ret %d\n", domain_string(alloc_domain), ret); @@ -1683,6 +1806,9 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu( pr_debug("Failed to allow vma node access. ret %d\n", ret); goto err_node_allow; } + ret = drm_gem_handle_create(adev->kfd.client.file, gobj, &(*mem)->gem_handle); + if (ret) + goto err_gem_handle_create; bo = gem_to_amdgpu_bo(gobj); if (bo_type == ttm_bo_type_sg) { bo->tbo.sg = sg; @@ -1697,6 +1823,7 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu( (*mem)->domain = domain; (*mem)->mapped_to_gpu_memory = 0; (*mem)->process_info = avm->process_info; + add_kgd_mem_to_kfd_bo_list(*mem, avm->process_info, user_addr); if (user_addr) { @@ -1713,6 +1840,15 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu( } bo->allowed_domains = AMDGPU_GEM_DOMAIN_GTT; bo->preferred_domains = AMDGPU_GEM_DOMAIN_GTT; + } else { + mutex_lock(&avm->process_info->lock); + if (avm->process_info->eviction_fence && + !dma_fence_is_signaled(&avm->process_info->eviction_fence->base)) + ret = amdgpu_amdkfd_bo_validate_and_fence(bo, domain, + &avm->process_info->eviction_fence->base); + mutex_unlock(&avm->process_info->lock); + if (ret) + goto err_validate_bo; } if (offset) @@ -1722,14 +1858,18 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu( allocate_init_user_pages_failed: err_pin_bo: +err_validate_bo: remove_kgd_mem_from_kfd_bo_list(*mem, avm->process_info); + drm_gem_handle_delete(adev->kfd.client.file, (*mem)->gem_handle); +err_gem_handle_create: drm_vma_node_revoke(&gobj->vma_node, drm_priv); err_node_allow: /* Don't unreserve system mem limit twice */ goto err_reserve_limit; err_bo_create: - amdgpu_amdkfd_unreserve_mem_limit(adev, aligned_size, flags); + amdgpu_amdkfd_unreserve_mem_limit(adev, aligned_size, flags, xcp_id); err_reserve_limit: + amdgpu_sync_free(&(*mem)->sync); mutex_destroy(&(*mem)->lock); if (gobj) drm_gem_object_put(gobj); @@ -1752,7 +1892,6 @@ int amdgpu_amdkfd_gpuvm_free_memory_of_gpu( bool use_release_notifier = (mem->bo->kfd_bo == mem); struct kfd_mem_attachment *entry, *tmp; struct bo_vm_reservation_context ctx; - struct ttm_validate_buffer *bo_list_entry; unsigned int mapped_to_gpu_memory; int ret; bool is_imported = false; @@ -1780,16 +1919,15 @@ int amdgpu_amdkfd_gpuvm_free_memory_of_gpu( } /* Make sure restore workers don't access the BO any more */ - bo_list_entry = &mem->validate_list; mutex_lock(&process_info->lock); - list_del(&bo_list_entry->head); + list_del(&mem->validate_list); mutex_unlock(&process_info->lock); /* Cleanup user pages and MMU notifiers */ if (amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm)) { amdgpu_hmm_unregister(mem->bo); mutex_lock(&process_info->notifier_lock); - amdgpu_ttm_tt_discard_user_pages(mem->bo->tbo.ttm, mem->range); + amdgpu_hmm_range_free(mem->range); mutex_unlock(&process_info->notifier_lock); } @@ -1797,18 +1935,16 @@ int amdgpu_amdkfd_gpuvm_free_memory_of_gpu( if (unlikely(ret)) return ret; - /* The eviction fence should be removed by the last unmap. - * TODO: Log an error condition if the bo still has the eviction fence - * attached - */ amdgpu_amdkfd_remove_eviction_fence(mem->bo, process_info->eviction_fence); pr_debug("Release VA 0x%llx - 0x%llx\n", mem->va, mem->va + bo_size * (1 + mem->aql_queue)); /* Remove from VM internal data structures */ - list_for_each_entry_safe(entry, tmp, &mem->attachments, list) + list_for_each_entry_safe(entry, tmp, &mem->attachments, list) { + kfd_mem_dmaunmap_attachment(mem, entry); kfd_mem_detach(entry); + } ret = unreserve_bo_and_vms(&ctx, false, false); @@ -1824,11 +1960,12 @@ int amdgpu_amdkfd_gpuvm_free_memory_of_gpu( } /* Update the size of the BO being freed if it was allocated from - * VRAM and is not imported. + * VRAM and is not imported. For APP APU VRAM allocations are done + * in GTT domain */ if (size) { - if ((mem->bo->preferred_domains == AMDGPU_GEM_DOMAIN_VRAM) && - (!is_imported)) + if (!is_imported && + mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) *size = bo_size; else *size = 0; @@ -1836,8 +1973,11 @@ int amdgpu_amdkfd_gpuvm_free_memory_of_gpu( /* Free the BO*/ drm_vma_node_revoke(&mem->bo->tbo.base.vma_node, drm_priv); - if (mem->dmabuf) + drm_gem_handle_delete(adev->kfd.client.file, mem->gem_handle); + if (mem->dmabuf) { dma_buf_put(mem->dmabuf); + mem->dmabuf = NULL; + } mutex_destroy(&mem->lock); /* If this releases the last reference, it will end up calling @@ -1920,23 +2060,10 @@ int amdgpu_amdkfd_gpuvm_map_memory_to_gpu( bo->tbo.resource->mem_type == TTM_PL_SYSTEM) is_invalid_userptr = true; - ret = vm_validate_pt_pd_bos(avm); + ret = vm_validate_pt_pd_bos(avm, NULL); if (unlikely(ret)) goto out_unreserve; - if (mem->mapped_to_gpu_memory == 0 && - !amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) { - /* Validate BO only once. The eviction fence gets added to BO - * the first time it is mapped. Validate will wait for all - * background evictions to complete. - */ - ret = amdgpu_amdkfd_bo_validate(bo, domain, true); - if (ret) { - pr_debug("Validate failed\n"); - goto out_unreserve; - } - } - list_for_each_entry(entry, &mem->attachments, list) { if (entry->bo_va->base.vm != avm || entry->is_mapped) continue; @@ -1963,10 +2090,6 @@ int amdgpu_amdkfd_gpuvm_map_memory_to_gpu( mem->mapped_to_gpu_memory); } - if (!amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) && !bo->tbo.pin_count) - dma_resv_add_fence(bo->tbo.base.resv, - &avm->process_info->eviction_fence->base, - DMA_RESV_USAGE_BOOKKEEP); ret = unreserve_bo_and_vms(&ctx, false, false); goto out; @@ -1979,11 +2102,41 @@ out: return ret; } +int amdgpu_amdkfd_gpuvm_dmaunmap_mem(struct kgd_mem *mem, void *drm_priv) +{ + struct kfd_mem_attachment *entry; + struct amdgpu_vm *vm; + int ret; + + vm = drm_priv_to_vm(drm_priv); + + mutex_lock(&mem->lock); + + ret = amdgpu_bo_reserve(mem->bo, true); + if (ret) + goto out; + + list_for_each_entry(entry, &mem->attachments, list) { + if (entry->bo_va->base.vm != vm) + continue; + if (entry->bo_va->base.bo->tbo.ttm && + !entry->bo_va->base.bo->tbo.ttm->sg) + continue; + + kfd_mem_dmaunmap_attachment(mem, entry); + } + + amdgpu_bo_unreserve(mem->bo); +out: + mutex_unlock(&mem->lock); + + return ret; +} + int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu( struct amdgpu_device *adev, struct kgd_mem *mem, void *drm_priv) { struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv); - struct amdkfd_process_info *process_info = avm->process_info; unsigned long bo_size = mem->bo->tbo.base.size; struct kfd_mem_attachment *entry; struct bo_vm_reservation_context ctx; @@ -2000,7 +2153,7 @@ int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu( goto unreserve_out; } - ret = vm_validate_pt_pd_bos(avm); + ret = vm_validate_pt_pd_bos(avm, NULL); if (unlikely(ret)) goto unreserve_out; @@ -2016,7 +2169,10 @@ int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu( pr_debug("\t unmap VA 0x%llx - 0x%llx from entry %p\n", entry->va, entry->va + bo_size, entry); - unmap_bo_from_gpuvm(mem, entry, ctx.sync); + ret = unmap_bo_from_gpuvm(mem, entry, ctx.sync); + if (ret) + goto unreserve_out; + entry->is_mapped = false; mem->mapped_to_gpu_memory--; @@ -2024,15 +2180,6 @@ int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu( mem->mapped_to_gpu_memory); } - /* If BO is unmapped from all VMs, unfence it. It can be evicted if - * required. - */ - if (mem->mapped_to_gpu_memory == 0 && - !amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm) && - !mem->bo->tbo.pin_count) - amdgpu_amdkfd_remove_eviction_fence(mem->bo, - process_info->eviction_fence); - unreserve_out: unreserve_bo_and_vms(&ctx, false, false); out: @@ -2059,13 +2206,13 @@ int amdgpu_amdkfd_gpuvm_sync_memory( /** * amdgpu_amdkfd_map_gtt_bo_to_gart - Map BO to GART and increment reference count - * @adev: Device to which allocated BO belongs * @bo: Buffer object to be mapped + * @bo_gart: Return bo reference * * Before return, bo reference count is incremented. To release the reference and unpin/ * unmap the BO, call amdgpu_amdkfd_free_gtt_mem. */ -int amdgpu_amdkfd_map_gtt_bo_to_gart(struct amdgpu_device *adev, struct amdgpu_bo *bo) +int amdgpu_amdkfd_map_gtt_bo_to_gart(struct amdgpu_bo *bo, struct amdgpu_bo **bo_gart) { int ret; @@ -2092,7 +2239,7 @@ int amdgpu_amdkfd_map_gtt_bo_to_gart(struct amdgpu_device *adev, struct amdgpu_b amdgpu_bo_unreserve(bo); - bo = amdgpu_bo_ref(bo); + *bo_gart = amdgpu_bo_ref(bo); return 0; @@ -2182,7 +2329,7 @@ void amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel(struct kgd_mem *mem) { struct amdgpu_bo *bo = mem->bo; - amdgpu_bo_reserve(bo, true); + (void)amdgpu_bo_reserve(bo, true); amdgpu_bo_kunmap(bo); amdgpu_bo_unpin(bo); amdgpu_bo_unreserve(bo); @@ -2191,34 +2338,24 @@ void amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel(struct kgd_mem *mem) int amdgpu_amdkfd_gpuvm_get_vm_fault_info(struct amdgpu_device *adev, struct kfd_vm_fault_info *mem) { - if (atomic_read(&adev->gmc.vm_fault_info_updated) == 1) { + if (atomic_read_acquire(&adev->gmc.vm_fault_info_updated) == 1) { *mem = *adev->gmc.vm_fault_info; - mb(); /* make sure read happened */ - atomic_set(&adev->gmc.vm_fault_info_updated, 0); + atomic_set_release(&adev->gmc.vm_fault_info_updated, 0); } return 0; } -int amdgpu_amdkfd_gpuvm_import_dmabuf(struct amdgpu_device *adev, - struct dma_buf *dma_buf, - uint64_t va, void *drm_priv, - struct kgd_mem **mem, uint64_t *size, - uint64_t *mmap_offset) +static int import_obj_create(struct amdgpu_device *adev, + struct dma_buf *dma_buf, + struct drm_gem_object *obj, + uint64_t va, void *drm_priv, + struct kgd_mem **mem, uint64_t *size, + uint64_t *mmap_offset) { struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv); - struct drm_gem_object *obj; struct amdgpu_bo *bo; int ret; - if (dma_buf->ops != &amdgpu_dmabuf_ops) - /* Can't handle non-graphics buffers */ - return -EINVAL; - - obj = dma_buf->priv; - if (drm_to_adev(obj->dev) != adev) - /* Can't handle buffers from other devices */ - return -EINVAL; - bo = gem_to_amdgpu_bo(obj); if (!(bo->preferred_domains & (AMDGPU_GEM_DOMAIN_VRAM | AMDGPU_GEM_DOMAIN_GTT))) @@ -2230,10 +2367,8 @@ int amdgpu_amdkfd_gpuvm_import_dmabuf(struct amdgpu_device *adev, return -ENOMEM; ret = drm_vma_node_allow(&obj->vma_node, drm_priv); - if (ret) { - kfree(*mem); - return ret; - } + if (ret) + goto err_free_mem; if (size) *size = amdgpu_bo_size(bo); @@ -2250,18 +2385,89 @@ int amdgpu_amdkfd_gpuvm_import_dmabuf(struct amdgpu_device *adev, | KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE | KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE; - drm_gem_object_get(&bo->tbo.base); + get_dma_buf(dma_buf); + (*mem)->dmabuf = dma_buf; (*mem)->bo = bo; (*mem)->va = va; - (*mem)->domain = (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) ? - AMDGPU_GEM_DOMAIN_VRAM : AMDGPU_GEM_DOMAIN_GTT; + (*mem)->domain = (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) && + !adev->apu_prefer_gtt ? + AMDGPU_GEM_DOMAIN_VRAM : AMDGPU_GEM_DOMAIN_GTT; + (*mem)->mapped_to_gpu_memory = 0; (*mem)->process_info = avm->process_info; add_kgd_mem_to_kfd_bo_list(*mem, avm->process_info, false); amdgpu_sync_create(&(*mem)->sync); (*mem)->is_imported = true; + mutex_lock(&avm->process_info->lock); + if (avm->process_info->eviction_fence && + !dma_fence_is_signaled(&avm->process_info->eviction_fence->base)) + ret = amdgpu_amdkfd_bo_validate_and_fence(bo, (*mem)->domain, + &avm->process_info->eviction_fence->base); + mutex_unlock(&avm->process_info->lock); + if (ret) + goto err_remove_mem; + + return 0; + +err_remove_mem: + remove_kgd_mem_from_kfd_bo_list(*mem, avm->process_info); + drm_vma_node_revoke(&obj->vma_node, drm_priv); +err_free_mem: + kfree(*mem); + return ret; +} + +int amdgpu_amdkfd_gpuvm_import_dmabuf_fd(struct amdgpu_device *adev, int fd, + uint64_t va, void *drm_priv, + struct kgd_mem **mem, uint64_t *size, + uint64_t *mmap_offset) +{ + struct drm_gem_object *obj; + uint32_t handle; + int ret; + + ret = drm_gem_prime_fd_to_handle(&adev->ddev, adev->kfd.client.file, fd, + &handle); + if (ret) + return ret; + obj = drm_gem_object_lookup(adev->kfd.client.file, handle); + if (!obj) { + ret = -EINVAL; + goto err_release_handle; + } + + ret = import_obj_create(adev, obj->dma_buf, obj, va, drm_priv, mem, size, + mmap_offset); + if (ret) + goto err_put_obj; + + (*mem)->gem_handle = handle; + return 0; + +err_put_obj: + drm_gem_object_put(obj); +err_release_handle: + drm_gem_handle_delete(adev->kfd.client.file, handle); + return ret; +} + +int amdgpu_amdkfd_gpuvm_export_dmabuf(struct kgd_mem *mem, + struct dma_buf **dma_buf) +{ + int ret; + + mutex_lock(&mem->lock); + ret = kfd_mem_export_dmabuf(mem); + if (ret) + goto out; + + get_dma_buf(mem->dmabuf); + *dma_buf = mem->dmabuf; +out: + mutex_unlock(&mem->lock); + return ret; } /* Evict a userptr BO by stopping the queues if necessary @@ -2294,10 +2500,14 @@ int amdgpu_amdkfd_evict_userptr(struct mmu_interval_notifier *mni, /* First eviction, stop the queues */ r = kgd2kfd_quiesce_mm(mni->mm, KFD_QUEUE_EVICTION_TRIGGER_USERPTR); - if (r) + + if (r && r != -ESRCH) pr_err("Failed to quiesce KFD\n"); - schedule_delayed_work(&process_info->restore_userptr_work, - msecs_to_jiffies(AMDGPU_USERPTR_RESTORE_DELAY_MS)); + + if (r != -ESRCH) + queue_delayed_work(system_freezable_wq, + &process_info->restore_userptr_work, + msecs_to_jiffies(AMDGPU_USERPTR_RESTORE_DELAY_MS)); } mutex_unlock(&process_info->notifier_lock); @@ -2324,14 +2534,14 @@ static int update_invalid_user_pages(struct amdkfd_process_info *process_info, /* Move all invalidated BOs to the userptr_inval_list */ list_for_each_entry_safe(mem, tmp_mem, &process_info->userptr_valid_list, - validate_list.head) + validate_list) if (mem->invalid) - list_move_tail(&mem->validate_list.head, + list_move_tail(&mem->validate_list, &process_info->userptr_inval_list); /* Go through userptr_inval_list and update any invalid user_pages */ list_for_each_entry(mem, &process_info->userptr_inval_list, - validate_list.head) { + validate_list) { invalid = mem->invalid; if (!invalid) /* BO hasn't been invalidated since the last @@ -2341,7 +2551,7 @@ static int update_invalid_user_pages(struct amdkfd_process_info *process_info, bo = mem->bo; - amdgpu_ttm_tt_discard_user_pages(bo->tbo.ttm, mem->range); + amdgpu_hmm_range_free(mem->range); mem->range = NULL; /* BO reservations and getting user pages (hmm_range_fault) @@ -2365,10 +2575,14 @@ static int update_invalid_user_pages(struct amdkfd_process_info *process_info, } } + mem->range = amdgpu_hmm_range_alloc(NULL); + if (unlikely(!mem->range)) + return -ENOMEM; /* Get updated user pages */ - ret = amdgpu_ttm_tt_get_user_pages(bo, bo->tbo.ttm->pages, - &mem->range); + ret = amdgpu_ttm_tt_get_user_pages(bo, mem->range); if (ret) { + amdgpu_hmm_range_free(mem->range); + mem->range = NULL; pr_debug("Failed %d to get user pages\n", ret); /* Return -EFAULT bad address error as success. It will @@ -2381,9 +2595,28 @@ static int update_invalid_user_pages(struct amdkfd_process_info *process_info, if (ret != -EFAULT) return ret; + /* If applications unmap memory before destroying the userptr + * from the KFD, trigger a segmentation fault in VM debug mode. + */ + if (amdgpu_ttm_adev(bo->tbo.bdev)->debug_vm_userptr) { + struct kfd_process *p; + + pr_err("Pid %d unmapped memory before destroying userptr at GPU addr 0x%llx\n", + pid_nr(process_info->pid), mem->va); + + // Send GPU VM fault to user space + p = kfd_lookup_process_by_pid(process_info->pid); + if (p) { + kfd_signal_vm_fault_event_with_userptr(p, mem->va); + kfd_unref_process(p); + } + } + ret = 0; } + amdgpu_ttm_tt_set_user_pages(bo->tbo.ttm, mem->range); + mutex_lock(&process_info->notifier_lock); /* Mark the BO as valid unless it was invalidated @@ -2393,7 +2626,9 @@ static int update_invalid_user_pages(struct amdkfd_process_info *process_info, ret = -EAGAIN; goto unlock_out; } - mem->invalid = 0; + /* set mem valid if mem has hmm range associated */ + if (mem->range) + mem->invalid = 0; } unlock_out: @@ -2409,59 +2644,50 @@ unlock_out: */ static int validate_invalid_user_pages(struct amdkfd_process_info *process_info) { - struct amdgpu_bo_list_entry *pd_bo_list_entries; - struct list_head resv_list, duplicates; - struct ww_acquire_ctx ticket; + struct ttm_operation_ctx ctx = { false, false }; struct amdgpu_sync sync; + struct drm_exec exec; struct amdgpu_vm *peer_vm; struct kgd_mem *mem, *tmp_mem; struct amdgpu_bo *bo; - struct ttm_operation_ctx ctx = { false, false }; - int i, ret; - - pd_bo_list_entries = kcalloc(process_info->n_vms, - sizeof(struct amdgpu_bo_list_entry), - GFP_KERNEL); - if (!pd_bo_list_entries) { - pr_err("%s: Failed to allocate PD BO list entries\n", __func__); - ret = -ENOMEM; - goto out_no_mem; - } - - INIT_LIST_HEAD(&resv_list); - INIT_LIST_HEAD(&duplicates); + int ret; - /* Get all the page directory BOs that need to be reserved */ - i = 0; - list_for_each_entry(peer_vm, &process_info->vm_list_head, - vm_list_node) - amdgpu_vm_get_pd_bo(peer_vm, &resv_list, - &pd_bo_list_entries[i++]); - /* Add the userptr_inval_list entries to resv_list */ - list_for_each_entry(mem, &process_info->userptr_inval_list, - validate_list.head) { - list_add_tail(&mem->resv_list.head, &resv_list); - mem->resv_list.bo = mem->validate_list.bo; - mem->resv_list.num_shared = mem->validate_list.num_shared; - } + amdgpu_sync_create(&sync); + drm_exec_init(&exec, 0, 0); /* Reserve all BOs and page tables for validation */ - ret = ttm_eu_reserve_buffers(&ticket, &resv_list, false, &duplicates); - WARN(!list_empty(&duplicates), "Duplicates should be empty"); - if (ret) - goto out_free; + drm_exec_until_all_locked(&exec) { + /* Reserve all the page directories */ + list_for_each_entry(peer_vm, &process_info->vm_list_head, + vm_list_node) { + ret = amdgpu_vm_lock_pd(peer_vm, &exec, 2); + drm_exec_retry_on_contention(&exec); + if (unlikely(ret)) + goto unreserve_out; + } - amdgpu_sync_create(&sync); + /* Reserve the userptr_inval_list entries to resv_list */ + list_for_each_entry(mem, &process_info->userptr_inval_list, + validate_list) { + struct drm_gem_object *gobj; + + gobj = &mem->bo->tbo.base; + ret = drm_exec_prepare_obj(&exec, gobj, 1); + drm_exec_retry_on_contention(&exec); + if (unlikely(ret)) + goto unreserve_out; + } + } - ret = process_validate_vms(process_info); + ret = process_validate_vms(process_info, NULL); if (ret) goto unreserve_out; /* Validate BOs and update GPUVM page tables */ list_for_each_entry_safe(mem, tmp_mem, &process_info->userptr_inval_list, - validate_list.head) { + validate_list) { struct kfd_mem_attachment *attachment; bo = mem->bo; @@ -2503,12 +2729,9 @@ static int validate_invalid_user_pages(struct amdkfd_process_info *process_info) ret = process_update_pds(process_info, &sync); unreserve_out: - ttm_eu_backoff_reservation(&ticket, &resv_list); + drm_exec_fini(&exec); amdgpu_sync_wait(&sync, false); amdgpu_sync_free(&sync); -out_free: - kfree(pd_bo_list_entries); -out_no_mem: return ret; } @@ -2524,9 +2747,16 @@ static int confirm_valid_user_pages_locked(struct amdkfd_process_info *process_i list_for_each_entry_safe(mem, tmp_mem, &process_info->userptr_inval_list, - validate_list.head) { - bool valid = amdgpu_ttm_tt_get_user_pages_done( - mem->bo->tbo.ttm, mem->range); + validate_list) { + bool valid; + + /* keep mem without hmm range at userptr_inval_list */ + if (!mem->range) + continue; + + /* Only check mem with hmm range associated */ + valid = amdgpu_hmm_range_valid(mem->range); + amdgpu_hmm_range_free(mem->range); mem->range = NULL; if (!valid) { @@ -2534,9 +2764,14 @@ static int confirm_valid_user_pages_locked(struct amdkfd_process_info *process_i ret = -EAGAIN; continue; } - WARN(mem->invalid, "Valid BO is marked invalid"); - list_move_tail(&mem->validate_list.head, + if (mem->invalid) { + WARN(1, "Valid BO is marked invalid"); + ret = -EAGAIN; + continue; + } + + list_move_tail(&mem->validate_list, &process_info->userptr_valid_list); } @@ -2617,7 +2852,8 @@ unlock_out: /* If validation failed, reschedule another attempt */ if (evicted_bos) { - schedule_delayed_work(&process_info->restore_userptr_work, + queue_delayed_work(system_freezable_wq, + &process_info->restore_userptr_work, msecs_to_jiffies(AMDGPU_USERPTR_RESTORE_DELAY_MS)); kfd_smi_event_queue_restore_rescheduled(mm); @@ -2626,6 +2862,23 @@ unlock_out: put_task_struct(usertask); } +static void replace_eviction_fence(struct dma_fence __rcu **ef, + struct dma_fence *new_ef) +{ + struct dma_fence *old_ef = rcu_replace_pointer(*ef, new_ef, true + /* protected by process_info->lock */); + + /* If we're replacing an unsignaled eviction fence, that fence will + * never be signaled, and if anyone is still waiting on that fence, + * they will hang forever. This should never happen. We should only + * replace the fence in restore_work that only gets scheduled after + * eviction work signaled the fence. + */ + WARN_ONCE(!dma_fence_is_signaled(old_ef), + "Replacing unsignaled eviction fence"); + dma_fence_put(old_ef); +} + /** amdgpu_amdkfd_gpuvm_restore_process_bos - Restore all BOs for the given * KFD process identified by process_info * @@ -2644,74 +2897,59 @@ unlock_out: * 7. Add fence to all PD and PT BOs. * 8. Unreserve all BOs */ -int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence **ef) +int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence __rcu **ef) { - struct amdgpu_bo_list_entry *pd_bo_list; struct amdkfd_process_info *process_info = info; struct amdgpu_vm *peer_vm; struct kgd_mem *mem; - struct bo_vm_reservation_context ctx; - struct amdgpu_amdkfd_fence *new_fence; - int ret = 0, i; struct list_head duplicate_save; struct amdgpu_sync sync_obj; unsigned long failed_size = 0; unsigned long total_size = 0; + struct drm_exec exec; + int ret; INIT_LIST_HEAD(&duplicate_save); - INIT_LIST_HEAD(&ctx.list); - INIT_LIST_HEAD(&ctx.duplicates); - - pd_bo_list = kcalloc(process_info->n_vms, - sizeof(struct amdgpu_bo_list_entry), - GFP_KERNEL); - if (!pd_bo_list) - return -ENOMEM; - i = 0; mutex_lock(&process_info->lock); - list_for_each_entry(peer_vm, &process_info->vm_list_head, - vm_list_node) - amdgpu_vm_get_pd_bo(peer_vm, &ctx.list, &pd_bo_list[i++]); - - /* Reserve all BOs and page tables/directory. Add all BOs from - * kfd_bo_list to ctx.list - */ - list_for_each_entry(mem, &process_info->kfd_bo_list, - validate_list.head) { - list_add_tail(&mem->resv_list.head, &ctx.list); - mem->resv_list.bo = mem->validate_list.bo; - mem->resv_list.num_shared = mem->validate_list.num_shared; - } + drm_exec_init(&exec, DRM_EXEC_IGNORE_DUPLICATES, 0); + drm_exec_until_all_locked(&exec) { + list_for_each_entry(peer_vm, &process_info->vm_list_head, + vm_list_node) { + ret = amdgpu_vm_lock_pd(peer_vm, &exec, 2); + drm_exec_retry_on_contention(&exec); + if (unlikely(ret)) { + pr_err("Locking VM PD failed, ret: %d\n", ret); + goto ttm_reserve_fail; + } + } - ret = ttm_eu_reserve_buffers(&ctx.ticket, &ctx.list, - false, &duplicate_save); - if (ret) { - pr_debug("Memory eviction: TTM Reserve Failed. Try again\n"); - goto ttm_reserve_fail; + /* Reserve all BOs and page tables/directory. Add all BOs from + * kfd_bo_list to ctx.list + */ + list_for_each_entry(mem, &process_info->kfd_bo_list, + validate_list) { + struct drm_gem_object *gobj; + + gobj = &mem->bo->tbo.base; + ret = drm_exec_prepare_obj(&exec, gobj, 1); + drm_exec_retry_on_contention(&exec); + if (unlikely(ret)) { + pr_err("drm_exec_prepare_obj failed, ret: %d\n", ret); + goto ttm_reserve_fail; + } + } } amdgpu_sync_create(&sync_obj); - /* Validate PDs and PTs */ - ret = process_validate_vms(process_info); - if (ret) - goto validate_map_fail; - - ret = process_sync_pds_resv(process_info, &sync_obj); - if (ret) { - pr_debug("Memory eviction: Failed to sync to PD BO moving fence. Try again\n"); - goto validate_map_fail; - } - - /* Validate BOs and map them to GPUVM (update VM page tables). */ + /* Validate BOs managed by KFD */ list_for_each_entry(mem, &process_info->kfd_bo_list, - validate_list.head) { + validate_list) { struct amdgpu_bo *bo = mem->bo; uint32_t domain = mem->domain; - struct kfd_mem_attachment *attachment; struct dma_resv_iter cursor; struct dma_fence *fence; @@ -2730,12 +2968,31 @@ int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence **ef) } dma_resv_for_each_fence(&cursor, bo->tbo.base.resv, DMA_RESV_USAGE_KERNEL, fence) { - ret = amdgpu_sync_fence(&sync_obj, fence); + ret = amdgpu_sync_fence(&sync_obj, fence, GFP_KERNEL); if (ret) { pr_debug("Memory eviction: Sync BO fence failed. Try again\n"); goto validate_map_fail; } } + } + + if (failed_size) + pr_debug("0x%lx/0x%lx in system\n", failed_size, total_size); + + /* Validate PDs, PTs and evicted DMABuf imports last. Otherwise BO + * validations above would invalidate DMABuf imports again. + */ + ret = process_validate_vms(process_info, &exec.ticket); + if (ret) { + pr_debug("Validating VMs failed, ret: %d\n", ret); + goto validate_map_fail; + } + + /* Update mappings managed by KFD. */ + list_for_each_entry(mem, &process_info->kfd_bo_list, + validate_list) { + struct kfd_mem_attachment *attachment; + list_for_each_entry(attachment, &mem->attachments, list) { if (!attachment->is_mapped) continue; @@ -2749,8 +3006,31 @@ int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence **ef) } } - if (failed_size) - pr_debug("0x%lx/0x%lx in system\n", failed_size, total_size); + /* Update mappings not managed by KFD */ + list_for_each_entry(peer_vm, &process_info->vm_list_head, + vm_list_node) { + struct amdgpu_device *adev = amdgpu_ttm_adev( + peer_vm->root.bo->tbo.bdev); + + struct amdgpu_fpriv *fpriv = + container_of(peer_vm, struct amdgpu_fpriv, vm); + + ret = amdgpu_vm_bo_update(adev, fpriv->prt_va, false); + if (ret) { + dev_dbg(adev->dev, + "Memory eviction: handle PRT moved failed, pid %8d. Try again.\n", + pid_nr(process_info->pid)); + goto validate_map_fail; + } + + ret = amdgpu_vm_handle_moved(adev, peer_vm, &exec.ticket); + if (ret) { + dev_dbg(adev->dev, + "Memory eviction: handle moved failed, pid %8d. Try again.\n", + pid_nr(process_info->pid)); + goto validate_map_fail; + } + } /* Update page directories */ ret = process_update_pds(process_info, &sync_obj); @@ -2759,29 +3039,50 @@ int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence **ef) goto validate_map_fail; } + /* Sync with fences on all the page tables. They implicitly depend on any + * move fences from amdgpu_vm_handle_moved above. + */ + ret = process_sync_pds_resv(process_info, &sync_obj); + if (ret) { + pr_debug("Memory eviction: Failed to sync to PD BO moving fence. Try again\n"); + goto validate_map_fail; + } + /* Wait for validate and PT updates to finish */ amdgpu_sync_wait(&sync_obj, false); - /* Release old eviction fence and create new one, because fence only - * goes from unsignaled to signaled, fence cannot be reused. - * Use context and mm from the old fence. + /* The old eviction fence may be unsignaled if restore happens + * after a GPU reset or suspend/resume. Keep the old fence in that + * case. Otherwise release the old eviction fence and create new + * one, because fence only goes from unsignaled to signaled once + * and cannot be reused. Use context and mm from the old fence. + * + * If an old eviction fence signals after this check, that's OK. + * Anyone signaling an eviction fence must stop the queues first + * and schedule another restore worker. */ - new_fence = amdgpu_amdkfd_fence_create( + if (dma_fence_is_signaled(&process_info->eviction_fence->base)) { + struct amdgpu_amdkfd_fence *new_fence = + amdgpu_amdkfd_fence_create( process_info->eviction_fence->base.context, process_info->eviction_fence->mm, NULL); - if (!new_fence) { - pr_err("Failed to create eviction fence\n"); - ret = -ENOMEM; - goto validate_map_fail; + + if (!new_fence) { + pr_err("Failed to create eviction fence\n"); + ret = -ENOMEM; + goto validate_map_fail; + } + dma_fence_put(&process_info->eviction_fence->base); + process_info->eviction_fence = new_fence; + replace_eviction_fence(ef, dma_fence_get(&new_fence->base)); + } else { + WARN_ONCE(*ef != &process_info->eviction_fence->base, + "KFD eviction fence doesn't match KGD process_info"); } - dma_fence_put(&process_info->eviction_fence->base); - process_info->eviction_fence = new_fence; - *ef = dma_fence_get(&new_fence->base); /* Attach new eviction fence to all BOs except pinned ones */ - list_for_each_entry(mem, &process_info->kfd_bo_list, - validate_list.head) { + list_for_each_entry(mem, &process_info->kfd_bo_list, validate_list) { if (mem->bo->tbo.pin_count) continue; @@ -2789,7 +3090,7 @@ int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence **ef) &process_info->eviction_fence->base, DMA_RESV_USAGE_BOOKKEEP); } - /* Attach eviction fence to PD / PT BOs */ + /* Attach eviction fence to PD / PT BOs and DMABuf imports */ list_for_each_entry(peer_vm, &process_info->vm_list_head, vm_list_node) { struct amdgpu_bo *bo = peer_vm->root.bo; @@ -2800,11 +3101,10 @@ int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence **ef) } validate_map_fail: - ttm_eu_backoff_reservation(&ctx.ticket, &ctx.list); amdgpu_sync_free(&sync_obj); ttm_reserve_fail: + drm_exec_fini(&exec); mutex_unlock(&process_info->lock); - kfree(pd_bo_list); return ret; } @@ -2920,12 +3220,13 @@ int amdgpu_amdkfd_get_tile_config(struct amdgpu_device *adev, return 0; } -bool amdgpu_amdkfd_bo_mapped_to_dev(struct amdgpu_device *adev, struct kgd_mem *mem) +bool amdgpu_amdkfd_bo_mapped_to_dev(void *drm_priv, struct kgd_mem *mem) { + struct amdgpu_vm *vm = drm_priv_to_vm(drm_priv); struct kfd_mem_attachment *entry; list_for_each_entry(entry, &mem->attachments, list) { - if (entry->is_mapped && entry->adev == adev) + if (entry->is_mapped && entry->bo_va->base.vm == vm) return true; } return false; |