[PATCH] drm/amdgpu: Add EXT_COHERENCE memory allocation flags
David Francis
David.Francis at amd.com
Mon Jul 24 14:17:02 UTC 2023
Cover letter got lost, here it is:
This is in support of a RCCL change that requires specific
coherence behaviour.
Corresponding Thunk patch is at
https://github.com/RadeonOpenCompute/ROCT-Thunk-Interface/pull/88
> These flags (for GEM and SVM allocations) allocate
> memory that allows for system-scope atomic semantics.
>
> On GFX943 these flags cause caches to be avoided on
> non-local memory.
>
> On all other ASICs they are identical in functionality to the
> equivalent COHERENT flags.
>
> Signed-off-by: David Francis<David.Francis at amd.com>
> ---
> drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 2 ++
> drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c | 1 +
> drivers/gpu/drm/amd/amdgpu/gmc_v11_0.c | 1 +
> drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c | 5 ++++-
> drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 10 +++++++++-
> include/uapi/drm/amdgpu_drm.h | 7 +++++++
> include/uapi/linux/kfd_ioctl.h | 3 +++
> 7 files changed, 27 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
> index 9e18fe5eb190..67634e9f6466 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
> @@ -1790,6 +1790,8 @@ 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_COHERENCE)
> + alloc_flags |= AMDGPU_GEM_CREATE_EXT_COHERENCE;
> if (flags & KFD_IOC_ALLOC_MEM_FLAGS_UNCACHED)
> alloc_flags |= AMDGPU_GEM_CREATE_UNCACHED;
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c b/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c
> index 1c07459e2bd2..6a6f6068bea0 100644
> --- a/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c
> +++ b/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c
> @@ -631,6 +631,7 @@ static void gmc_v10_0_get_vm_pte(struct amdgpu_device *adev,
> }
>
> if (bo && bo->flags & (AMDGPU_GEM_CREATE_COHERENT |
> + AMDGPU_GEM_CREATE_EXT_COHERENCE |
> AMDGPU_GEM_CREATE_UNCACHED))
> *flags = (*flags & ~AMDGPU_PTE_MTYPE_NV10_MASK) |
> AMDGPU_PTE_MTYPE_NV10(MTYPE_UC);
> diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v11_0.c b/drivers/gpu/drm/amd/amdgpu/gmc_v11_0.c
> index a6ee0220db56..ff330c7c0232 100644
> --- a/drivers/gpu/drm/amd/amdgpu/gmc_v11_0.c
> +++ b/drivers/gpu/drm/amd/amdgpu/gmc_v11_0.c
> @@ -540,6 +540,7 @@ static void gmc_v11_0_get_vm_pte(struct amdgpu_device *adev,
> }
>
> if (bo && bo->flags & (AMDGPU_GEM_CREATE_COHERENT |
> + AMDGPU_GEM_CREATE_EXT_COHERENCE |
> AMDGPU_GEM_CREATE_UNCACHED))
> *flags = (*flags & ~AMDGPU_PTE_MTYPE_NV10_MASK) |
> AMDGPU_PTE_MTYPE_NV10(MTYPE_UC);
> diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c b/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c
> index 8447fcada8bb..17cf19c868e1 100644
> --- a/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c
> +++ b/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c
> @@ -1180,7 +1180,8 @@ static void gmc_v9_0_get_coherence_flags(struct amdgpu_device *adev,
> {
> struct amdgpu_device *bo_adev = amdgpu_ttm_adev(bo->tbo.bdev);
> bool is_vram = bo->tbo.resource->mem_type == TTM_PL_VRAM;
> - bool coherent = bo->flags & AMDGPU_GEM_CREATE_COHERENT;
> + bool coherent = bo->flags & (AMDGPU_GEM_CREATE_COHERENT | AMDGPU_GEM_CREATE_EXT_COHERENCE);
> + bool ext_coherence = bo->flags & AMDGPU_GEM_CREATE_EXT_COHERENCE;
> bool uncached = bo->flags & AMDGPU_GEM_CREATE_UNCACHED;
> struct amdgpu_vm *vm = mapping->bo_va->base.vm;
> unsigned int mtype_local, mtype;
> @@ -1248,6 +1249,8 @@ static void gmc_v9_0_get_coherence_flags(struct amdgpu_device *adev,
> snoop = true;
> if (uncached) {
> mtype = MTYPE_UC;
> + } else if (ext_coherence) {
> + mtype = is_local ? MTYPE_CC : MTYPE_UC;
> } else if (adev->flags & AMD_IS_APU) {
> mtype = is_local ? mtype_local : MTYPE_NC;
> } else {
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
> index 909f1ef8927d..acb87b2fe8df 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
> @@ -1157,7 +1157,8 @@ svm_range_get_pte_flags(struct kfd_node *node,
> uint32_t mapping_flags = 0;
> uint64_t pte_flags;
> bool snoop = (domain != SVM_RANGE_VRAM_DOMAIN);
> - bool coherent = flags & KFD_IOCTL_SVM_FLAG_COHERENT;
> + bool coherent = flags & (KFD_IOCTL_SVM_FLAG_COHERENT | KFD_IOCTL_SVM_FLAG_EXT_COHERENCE);
> + bool ext_coherence = flags & KFD_IOCTL_SVM_FLAG_EXT_COHERENCE;
> bool uncached = false; /*flags & KFD_IOCTL_SVM_FLAG_UNCACHED;*/
> unsigned int mtype_local;
>
> @@ -1205,6 +1206,13 @@ svm_range_get_pte_flags(struct kfd_node *node,
> snoop = true;
> if (uncached) {
> mapping_flags |= AMDGPU_VM_MTYPE_UC;
> + } else if (ext_coherence) {
> + /* local HBM region close to partition */
> + if (bo_node->adev == node->adev &&
> + (!bo_node->xcp || !node->xcp || bo_node->xcp->mem_id == node->xcp->mem_id))
> + mapping_flags |= AMDGPU_VM_MTYPE_CC;
> + else
> + mapping_flags |= AMDGPU_VM_MTYPE_UC;
> } else if (domain == SVM_RANGE_VRAM_DOMAIN) {
> /* local HBM region close to partition */
> if (bo_node->adev == node->adev &&
> diff --git a/include/uapi/drm/amdgpu_drm.h b/include/uapi/drm/amdgpu_drm.h
> index 3b98635bdfa8..7dbee3e3d9d1 100644
> --- a/include/uapi/drm/amdgpu_drm.h
> +++ b/include/uapi/drm/amdgpu_drm.h
> @@ -173,6 +173,13 @@ extern "C" {
> * may override the MTYPE selected in AMDGPU_VA_OP_MAP.
> */
> #define AMDGPU_GEM_CREATE_UNCACHED (1 << 14)
> +/* Flag that BO should be coherent across devices when using device-level
> + * atomics. May depend on GPU instructions to flush caches explicitly.
> + *
> + * This influences the choice of MTYPE in the PTEs on GFXv9 and later GPUs and
> + * may override the MTYPE selected in AMDGPU_VA_OP_MAP.
> + */
> +#define AMDGPU_GEM_CREATE_EXT_COHERENCE (1 << 15)
>
> /* hybrid specific */
> /* Flag that the memory should be in SPARSE resource */
> diff --git a/include/uapi/linux/kfd_ioctl.h b/include/uapi/linux/kfd_ioctl.h
> index e467ac678ba9..7a4041ee5951 100644
> --- a/include/uapi/linux/kfd_ioctl.h
> +++ b/include/uapi/linux/kfd_ioctl.h
> @@ -416,6 +416,7 @@ struct kfd_ioctl_acquire_vm_args {
> #define KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM (1 << 27)
> #define KFD_IOC_ALLOC_MEM_FLAGS_COHERENT (1 << 26)
> #define KFD_IOC_ALLOC_MEM_FLAGS_UNCACHED (1 << 25)
> +#define KFD_IOC_ALLOC_MEM_FLAGS_EXT_COHERENCE (1 << 24)
>
> /* Allocate memory for later SVM (shared virtual memory) mapping.
> *
> @@ -777,6 +778,8 @@ struct kfd_ioctl_cross_memory_copy_deprecated_args {
> #define KFD_IOCTL_SVM_FLAG_GPU_READ_MOSTLY 0x00000020
> /* Keep GPU memory mapping always valid as if XNACK is disable */
> #define KFD_IOCTL_SVM_FLAG_GPU_ALWAYS_MAPPED 0x00000040
> +/* Fine grained coherency between all devices using device-scope atomics */
> +#define KFD_IOCTL_SVM_FLAG_EXT_COHERENCE 0x00000080
>
> /**
> * kfd_ioctl_svm_op - SVM ioctl operations
> -- 2.34.1
More information about the amd-gfx
mailing list