[PATCH v7 1/1] gpu/cuda: introduce CUDA driver

[<eagostini@nvidia.com>]

From: Elena Agostini <eagostini@nvidia.com>

This is the CUDA implementation of the gpudev library.
Functionalities implemented through CUDA Driver API are:
- Device probe and remove
- Manage device memory allocations
- Register/unregister external CPU memory in the device memory area

Signed-off-by: Elena Agostini <eagostini@nvidia.com>
---
 doc/guides/gpus/cuda.rst               |  152 ++++
 doc/guides/gpus/index.rst              |    1 +
 doc/guides/rel_notes/release_21_11.rst |    2 +
 drivers/gpu/cuda/cuda.c                | 1150 ++++++++++++++++++++++++
 drivers/gpu/cuda/meson.build           |   18 +
 drivers/gpu/cuda/version.map           |    3 +
 drivers/gpu/meson.build                |    2 +-
 7 files changed, 1327 insertions(+), 1 deletion(-)
 create mode 100644 doc/guides/gpus/cuda.rst
 create mode 100644 drivers/gpu/cuda/cuda.c
 create mode 100644 drivers/gpu/cuda/meson.build
 create mode 100644 drivers/gpu/cuda/version.map

diff --git a/doc/guides/gpus/cuda.rst b/doc/guides/gpus/cuda.rst
new file mode 100644
index 0000000000..d007c9ff2c
--- /dev/null
+++ b/doc/guides/gpus/cuda.rst
@@ -0,0 +1,152 @@
+.. SPDX-License-Identifier: BSD-3-Clause
+   Copyright (c) 2021 NVIDIA Corporation & Affiliates
+
+CUDA GPU driver
+===============
+
+The CUDA GPU driver library (**librte_gpu_cuda**) provides support for NVIDIA GPUs.
+Information and documentation about these devices can be found on the
+`NVIDIA website <http://www.nvidia.com>`_. Help is also provided by the
+`NVIDIA CUDA Toolkit developer zone <https://docs.nvidia.com/cuda>`_.
+
+Build dependencies
+------------------
+
+The CUDA GPU driver library has an header-only dependency on ``cuda.h`` and ``cudaTypedefs.h``.
+To get these headers there are two options:
+
+- Install `CUDA Toolkit <https://developer.nvidia.com/cuda-toolkit>`_ 
+  (either regular or stubs installation).
+- Download these two headers from this `CUDA headers
+  <https://gitlab.com/nvidia/headers/cuda-individual/cudart>`_ public repo.
+
+You need to indicate to meson where CUDA headers files are through the CFLAGS variable.
+Three ways:
+
+- Set ``export CFLAGS=-I/usr/local/cuda/include`` before building
+- Add CFLAGS in the meson command line ``CFLAGS=-I/usr/local/cuda/include meson build``
+- Add the ``-Dc_args`` in meson command line ``meson build -Dc_args=-I/usr/local/cuda/include``
+
+If headers are not found, the CUDA GPU driver library is not built.
+
+CUDA Shared Library
+-------------------
+
+To avoid any system configuration issue, the CUDA API **libcuda.so** shared library
+is not linked at building time because of a Meson bug that looks
+for `cudart` module even if the `meson.build` file only requires default `cuda` module.
+
+**libcuda.so** is loaded at runtime in the ``cuda_gpu_probe`` function through ``dlopen``
+when the very first GPU is detected.
+If CUDA installation resides in a custom directory,
+the environment variable ``CUDA_PATH_L`` should specify where ``dlopen``
+can look for **libcuda.so**.
+
+All CUDA API symbols are loaded at runtime as well.
+For this reason, to build the CUDA driver library,
+no need to install the CUDA library.
+
+Design
+------
+
+**librte_gpu_cuda** relies on CUDA Driver API (no need for CUDA Runtime API).
+
+Goal of this driver library is not to provide a wrapper for the whole CUDA Driver API.
+Instead, the scope is to implement the generic features of gpudev API.
+For a CUDA application, integrating the gpudev library functions
+using the CUDA driver library is quite straightforward
+and doesn't create any compatibility problem.
+
+Initialization
+~~~~~~~~~~~~~~
+
+During initialization, CUDA driver library detects NVIDIA physical GPUs
+on the system or specified via EAL device options (e.g. ``-a b6:00.0``).
+The driver initializes the CUDA driver environment through ``cuInit(0)`` function.
+For this reason, it's required to set any CUDA environment configuration before
+calling ``rte_eal_init`` function in the DPDK application.
+
+If the CUDA driver environment has been already initialized, the ``cuInit(0)``
+in CUDA driver library has no effect.
+
+CUDA Driver sub-contexts
+~~~~~~~~~~~~~~~~~~~~~~~~
+
+After initialization, a CUDA application can create multiple sub-contexts
+on GPU physical devices.
+Through gpudev library, is possible to register these sub-contexts
+in the CUDA driver library as child devices having as parent a GPU physical device.
+
+CUDA driver library also supports `MPS
+<https://docs.nvidia.com/deploy/pdf/CUDA_Multi_Process_Service_Overview.pdf>`__.
+
+GPU memory management
+~~~~~~~~~~~~~~~~~~~~~
+
+The CUDA driver library maintains a table of GPU memory addresses allocated
+and CPU memory addresses registered associated to the input CUDA context.
+Whenever the application tried to deallocate or deregister a memory address,
+if the address is not in the table the CUDA driver library will return an error.
+
+Features
+--------
+
+- Register new child devices aka new CUDA Driver contexts.
+- Allocate memory on the GPU.
+- Register CPU memory to make it visible from GPU.
+
+Minimal requirements
+--------------------
+
+Minimal requirements to enable the CUDA driver library are:
+
+- NVIDIA GPU Ampere or Volta
+- CUDA 11.4 Driver API or newer
+
+`GPUDirect RDMA Technology <https://docs.nvidia.com/cuda/gpudirect-rdma/index.html>`_
+allows compatible network cards (e.g. Mellanox) to directly send and receive packets
+using GPU memory instead of additional memory copies through the CPU system memory.
+To enable this technology, system requirements are:
+
+- `nvidia-peermem <https://docs.nvidia.com/cuda/gpudirect-rdma/index.html#nvidia-peermem>`_
+  module running on the system;
+- Mellanox network card ConnectX-5 or newer (BlueField models included);
+- DPDK mlx5 PMD enabled;
+- To reach the best performance, an additional PCIe switch between GPU and NIC is recommended.
+
+Limitations
+-----------
+
+Supported only on Linux.
+
+Supported GPUs
+--------------
+
+The following NVIDIA GPU devices are supported by this CUDA driver library:
+
+- NVIDIA A100 80GB PCIe
+- NVIDIA A100 40GB PCIe
+- NVIDIA A30 24GB
+- NVIDIA A10 24GB
+- NVIDIA V100 32GB PCIe
+- NVIDIA V100 16GB PCIe
+
+External references
+-------------------
+
+A good example of how to use the GPU CUDA driver library through the gpudev library
+is the l2fwd-nv application that can be found `here <https://github.com/NVIDIA/l2fwd-nv>`_.
+
+The application is based on vanilla DPDK example l2fwd
+and is enhanced with GPU memory managed through gpudev library
+and CUDA to launch the swap of packets MAC addresses workload on the GPU.
+
+l2fwd-nv is not intended to be used for performance
+(testpmd is the good candidate for this).
+The goal is to show different use-cases about how a CUDA application can use DPDK to:
+
+- Allocate memory on GPU device using gpudev library.
+- Use that memory to create an external GPU memory mempool.
+- Receive packets directly in GPU memory.
+- Coordinate the workload on the GPU with the network and CPU activity to receive packets.
+- Send modified packets directly from the GPU memory.
diff --git a/doc/guides/gpus/index.rst b/doc/guides/gpus/index.rst
index 1878423239..4b7a420556 100644
--- a/doc/guides/gpus/index.rst
+++ b/doc/guides/gpus/index.rst
@@ -9,3 +9,4 @@ General-Purpose Graphics Processing Unit Drivers
    :numbered:
 
    overview
+   cuda
diff --git a/doc/guides/rel_notes/release_21_11.rst b/doc/guides/rel_notes/release_21_11.rst
index cd4dcd0077..d76bba2fe3 100644
--- a/doc/guides/rel_notes/release_21_11.rst
+++ b/doc/guides/rel_notes/release_21_11.rst
@@ -111,6 +111,8 @@ New Features
   * Memory management
   * Communication flag & list
 
+* **Added NVIDIA GPU driver implemented with CUDA library.**
+
 * **Added new RSS offload types for IPv4/L4 checksum in RSS flow.**
 
   Added macros ETH_RSS_IPV4_CHKSUM and ETH_RSS_L4_CHKSUM, now IPv4 and
diff --git a/drivers/gpu/cuda/cuda.c b/drivers/gpu/cuda/cuda.c
new file mode 100644
index 0000000000..12217c48c6
--- /dev/null
+++ b/drivers/gpu/cuda/cuda.c
@@ -0,0 +1,1150 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright (c) 2021 NVIDIA Corporation & Affiliates
+ */
+
+#include <dlfcn.h>
+
+#include <rte_common.h>
+#include <rte_log.h>
+#include <rte_malloc.h>
+#include <rte_errno.h>
+#include <rte_pci.h>
+#include <rte_bus_pci.h>
+#include <rte_byteorder.h>
+#include <rte_dev.h>
+
+#include <gpudev_driver.h>
+#include <cuda.h>
+#include <cudaTypedefs.h>
+
+#define CUDA_DRIVER_MIN_VERSION 11040
+#define CUDA_API_MIN_VERSION 3020
+
+/* CUDA Driver functions loaded with dlsym() */
+CUresult CUDAAPI (*sym_cuInit)(unsigned int flags) = NULL;
+CUresult CUDAAPI (*sym_cuDriverGetVersion)(int *driverVersion) = NULL;
+CUresult CUDAAPI (*sym_cuGetProcAddress)(const char *symbol,
+		void **pfn, int cudaVersion, uint64_t flags) = NULL;
+
+/* CUDA Driver functions loaded with cuGetProcAddress for versioning */
+PFN_cuGetErrorString pfn_cuGetErrorString;
+PFN_cuGetErrorName pfn_cuGetErrorName;
+PFN_cuPointerSetAttribute pfn_cuPointerSetAttribute;
+PFN_cuDeviceGetAttribute pfn_cuDeviceGetAttribute;
+PFN_cuDeviceGetByPCIBusId pfn_cuDeviceGetByPCIBusId;
+PFN_cuDevicePrimaryCtxRetain pfn_cuDevicePrimaryCtxRetain;
+PFN_cuDevicePrimaryCtxRelease pfn_cuDevicePrimaryCtxRelease;
+PFN_cuDeviceTotalMem pfn_cuDeviceTotalMem;
+PFN_cuDeviceGetName pfn_cuDeviceGetName;
+PFN_cuCtxGetApiVersion pfn_cuCtxGetApiVersion;
+PFN_cuCtxSetCurrent pfn_cuCtxSetCurrent;
+PFN_cuCtxGetCurrent pfn_cuCtxGetCurrent;
+PFN_cuCtxGetDevice pfn_cuCtxGetDevice;
+PFN_cuCtxGetExecAffinity pfn_cuCtxGetExecAffinity;
+PFN_cuMemAlloc pfn_cuMemAlloc;
+PFN_cuMemFree pfn_cuMemFree;
+PFN_cuMemHostRegister pfn_cuMemHostRegister;
+PFN_cuMemHostUnregister pfn_cuMemHostUnregister;
+PFN_cuMemHostGetDevicePointer pfn_cuMemHostGetDevicePointer;
+PFN_cuFlushGPUDirectRDMAWrites pfn_cuFlushGPUDirectRDMAWrites;
+
+static void *cudalib;
+static unsigned int cuda_api_version;
+static int cuda_driver_version;
+
+/* NVIDIA GPU vendor */
+#define NVIDIA_GPU_VENDOR_ID (0x10de)
+
+/* NVIDIA GPU device IDs */
+#define NVIDIA_GPU_A100_40GB_DEVICE_ID (0x20f1)
+#define NVIDIA_GPU_A100_80GB_DEVICE_ID (0x20b5)
+
+#define NVIDIA_GPU_A30_24GB_DEVICE_ID (0x20b7)
+#define NVIDIA_GPU_A10_24GB_DEVICE_ID (0x2236)
+
+#define NVIDIA_GPU_V100_32GB_DEVICE_ID (0x1db6)
+#define NVIDIA_GPU_V100_16GB_DEVICE_ID (0x1db4)
+
+#define NVIDIA_GPU_T4_16GB_DEVICE_ID (0x1eb8)
+
+#define CUDA_MAX_ALLOCATION_NUM 512
+
+#define GPU_PAGE_SHIFT 16
+#define GPU_PAGE_SIZE (1UL << GPU_PAGE_SHIFT)
+
+static RTE_LOG_REGISTER_DEFAULT(cuda_logtype, NOTICE);
+
+/* Helper macro for logging */
+#define rte_cuda_log(level, fmt, ...) \
+	rte_log(RTE_LOG_ ## level, cuda_logtype, fmt "\n", ##__VA_ARGS__)
+
+#define rte_cuda_debug(fmt, ...) \
+	rte_cuda_log(DEBUG, RTE_STR(__LINE__) ":%s() " fmt, __func__, \
+		##__VA_ARGS__)
+
+/* NVIDIA GPU address map */
+static struct rte_pci_id pci_id_cuda_map[] = {
+	{
+		RTE_PCI_DEVICE(NVIDIA_GPU_VENDOR_ID,
+				NVIDIA_GPU_A100_40GB_DEVICE_ID)
+	},
+	{
+		RTE_PCI_DEVICE(NVIDIA_GPU_VENDOR_ID,
+				NVIDIA_GPU_A100_80GB_DEVICE_ID)
+	},
+	{
+		RTE_PCI_DEVICE(NVIDIA_GPU_VENDOR_ID,
+				NVIDIA_GPU_A30_24GB_DEVICE_ID)
+	},
+	{
+		RTE_PCI_DEVICE(NVIDIA_GPU_VENDOR_ID,
+				NVIDIA_GPU_A10_24GB_DEVICE_ID)
+	},
+	{
+		RTE_PCI_DEVICE(NVIDIA_GPU_VENDOR_ID,
+				NVIDIA_GPU_V100_32GB_DEVICE_ID)
+	},
+	{
+		RTE_PCI_DEVICE(NVIDIA_GPU_VENDOR_ID,
+				NVIDIA_GPU_V100_16GB_DEVICE_ID)
+	},
+	{
+		RTE_PCI_DEVICE(NVIDIA_GPU_VENDOR_ID,
+				NVIDIA_GPU_T4_16GB_DEVICE_ID)
+	},
+	{
+		.device_id = 0
+	}
+};
+
+/* Device private info */
+struct cuda_info {
+	char gpu_name[RTE_DEV_NAME_MAX_LEN];
+	CUdevice cu_dev;
+	int gdr_supported;
+	int gdr_write_ordering;
+	int gdr_flush_type;
+};
+
+/* Type of memory allocated by CUDA driver */
+enum mem_type {
+	GPU_MEM = 0,
+	CPU_REGISTERED,
+	GPU_REGISTERED /* Not used yet */
+};
+
+/* key associated to a memory address */
+typedef uintptr_t cuda_ptr_key;
+
+/* Single entry of the memory list */
+struct mem_entry {
+	CUdeviceptr ptr_d;
+	void *ptr_h;
+	size_t size;
+	struct rte_gpu *dev;
+	CUcontext ctx;
+	cuda_ptr_key pkey;
+	enum mem_type mtype;
+	struct mem_entry *prev;
+	struct mem_entry *next;
+};
+
+static struct mem_entry *mem_alloc_list_head;
+static struct mem_entry *mem_alloc_list_tail;
+static uint32_t mem_alloc_list_last_elem;
+
+/* Load the CUDA symbols */
+
+static int
+cuda_loader(void)
+{
+	char cuda_path[1024];
+
+	if (getenv("CUDA_PATH_L") == NULL)
+		snprintf(cuda_path, 1024, "%s", "libcuda.so");
+	else
+		snprintf(cuda_path, 1024, "%s%s", getenv("CUDA_PATH_L"), "libcuda.so");
+
+	cudalib = dlopen(cuda_path, RTLD_LAZY);
+	if (cudalib == NULL) {
+		rte_cuda_log(ERR, "Failed to find CUDA library in %s (CUDA_PATH_L=%s)",
+				cuda_path, getenv("CUDA_PATH_L"));
+		return -1;
+	}
+
+	return 0;
+}
+
+static int
+cuda_sym_func_loader(void)
+{
+	if (cudalib == NULL)
+		return -1;
+
+	sym_cuInit = dlsym(cudalib, "cuInit");
+	if (sym_cuInit == NULL) {
+		rte_cuda_log(ERR, "Failed to load CUDA missing symbol cuInit");
+		return -1;
+	}
+
+	sym_cuDriverGetVersion = dlsym(cudalib, "cuDriverGetVersion");
+	if (sym_cuDriverGetVersion == NULL) {
+		rte_cuda_log(ERR, "Failed to load CUDA missing symbol cuDriverGetVersion");
+		return -1;
+	}
+
+	sym_cuGetProcAddress = dlsym(cudalib, "cuGetProcAddress");
+	if (sym_cuGetProcAddress == NULL) {
+		rte_cuda_log(ERR, "Failed to load CUDA missing symbol cuGetProcAddress");
+		return -1;
+	}
+
+	return 0;
+}
+
+static int
+cuda_pfn_func_loader(void)
+{
+	CUresult res;
+
+	res = sym_cuGetProcAddress("cuGetErrorString",
+			(void **) (&pfn_cuGetErrorString), cuda_driver_version, 0);
+	if (res != 0) {
+		rte_cuda_log(ERR, "Retrieve pfn_cuGetErrorString failed with %d", res);
+		return -1;
+	}
+
+	res = sym_cuGetProcAddress("cuGetErrorName",
+			(void **) (&pfn_cuGetErrorName), cuda_driver_version, 0);
+	if (res != 0) {
+		rte_cuda_log(ERR, "Retrieve pfn_cuGetErrorName failed with %d", res);
+		return -1;
+	}
+
+	res = sym_cuGetProcAddress("cuPointerSetAttribute",
+			(void **) (&pfn_cuPointerSetAttribute), cuda_driver_version, 0);
+	if (res != 0) {
+		rte_cuda_log(ERR, "Retrieve pfn_cuPointerSetAttribute failed with %d", res);
+		return -1;
+	}
+
+	res = sym_cuGetProcAddress("cuDeviceGetAttribute",
+			(void **) (&pfn_cuDeviceGetAttribute), cuda_driver_version, 0);
+	if (res != 0) {
+		rte_cuda_log(ERR, "Retrieve pfn_cuDeviceGetAttribute failed with %d", res);
+		return -1;
+	}
+
+	res = sym_cuGetProcAddress("cuDeviceGetByPCIBusId",
+			(void **) (&pfn_cuDeviceGetByPCIBusId), cuda_driver_version, 0);
+	if (res != 0) {
+		rte_cuda_log(ERR, "Retrieve pfn_cuDeviceGetByPCIBusId failed with %d", res);
+		return -1;
+	}
+
+	res = sym_cuGetProcAddress("cuDeviceGetName",
+			(void **) (&pfn_cuDeviceGetName), cuda_driver_version, 0);
+	if (res != 0) {
+		rte_cuda_log(ERR, "Retrieve pfn_cuDeviceGetName failed with %d", res);
+		return -1;
+	}
+
+	res = sym_cuGetProcAddress("cuDevicePrimaryCtxRetain",
+			(void **) (&pfn_cuDevicePrimaryCtxRetain), cuda_driver_version, 0);
+	if (res != 0) {
+		rte_cuda_log(ERR, "Retrieve pfn_cuDevicePrimaryCtxRetain failed with %d", res);
+		return -1;
+	}
+
+	res = sym_cuGetProcAddress("cuDevicePrimaryCtxRelease",
+			(void **) (&pfn_cuDevicePrimaryCtxRelease), cuda_driver_version, 0);
+	if (res != 0) {
+		rte_cuda_log(ERR, "Retrieve pfn_cuDevicePrimaryCtxRelease failed with %d", res);
+		return -1;
+	}
+
+	res = sym_cuGetProcAddress("cuDeviceTotalMem",
+			(void **) (&pfn_cuDeviceTotalMem), cuda_driver_version, 0);
+	if (res != 0) {
+		rte_cuda_log(ERR, "Retrieve pfn_cuDeviceTotalMem failed with %d", res);
+		return -1;
+	}
+
+	res = sym_cuGetProcAddress("cuCtxGetApiVersion",
+			(void **) (&pfn_cuCtxGetApiVersion), cuda_driver_version, 0);
+	if (res != 0) {
+		rte_cuda_log(ERR, "Retrieve pfn_cuCtxGetApiVersion failed with %d", res);
+		return -1;
+	}
+
+	res = sym_cuGetProcAddress("cuCtxGetDevice",
+			(void **) (&pfn_cuCtxGetDevice), cuda_driver_version, 0);
+	if (res != 0) {
+		rte_cuda_log(ERR, "Retrieve pfn_cuCtxGetDevice failed with %d", res);
+		return -1;
+	}
+
+	res = sym_cuGetProcAddress("cuCtxSetCurrent",
+			(void **) (&pfn_cuCtxSetCurrent), cuda_driver_version, 0);
+	if (res != 0) {
+		rte_cuda_log(ERR, "Retrieve pfn_cuCtxSetCurrent failed with %d", res);
+		return -1;
+	}
+
+	res = sym_cuGetProcAddress("cuCtxGetCurrent",
+			(void **) (&pfn_cuCtxGetCurrent), cuda_driver_version, 0);
+	if (res != 0) {
+		rte_cuda_log(ERR, "Retrieve pfn_cuCtxGetCurrent failed with %d", res);
+		return -1;
+	}
+
+	res = sym_cuGetProcAddress("cuCtxGetExecAffinity",
+			(void **) (&pfn_cuCtxGetExecAffinity), cuda_driver_version, 0);
+	if (res != 0) {
+		rte_cuda_log(ERR, "Retrieve pfn_cuCtxGetExecAffinity failed with %d", res);
+		return -1;
+	}
+
+	res = sym_cuGetProcAddress("cuMemAlloc",
+			(void **) (&pfn_cuMemAlloc), cuda_driver_version, 0);
+	if (res != 0) {
+		rte_cuda_log(ERR, "Retrieve pfn_cuMemAlloc failed with %d", res);
+		return -1;
+	}
+
+	res = sym_cuGetProcAddress("cuMemFree",
+			(void **) (&pfn_cuMemFree), cuda_driver_version, 0);
+	if (res != 0) {
+		rte_cuda_log(ERR, "Retrieve pfn_cuMemFree failed with %d", res);
+		return -1;
+	}
+
+	res = sym_cuGetProcAddress("cuMemHostRegister",
+			(void **) (&pfn_cuMemHostRegister), cuda_driver_version, 0);
+	if (res != 0) {
+		rte_cuda_log(ERR, "Retrieve pfn_cuMemHostRegister failed with %d", res);
+		return -1;
+	}
+
+	res = sym_cuGetProcAddress("cuMemHostUnregister",
+			(void **) (&pfn_cuMemHostUnregister), cuda_driver_version, 0);
+	if (res != 0) {
+		rte_cuda_log(ERR, "Retrieve pfn_cuMemHostUnregister failed with %d", res);
+		return -1;
+	}
+
+	res = sym_cuGetProcAddress("cuMemHostGetDevicePointer",
+			(void **) (&pfn_cuMemHostGetDevicePointer), cuda_driver_version, 0);
+	if (res != 0) {
+		rte_cuda_log(ERR, "Retrieve pfn_cuMemHostGetDevicePointer failed with %d", res);
+		return -1;
+	}
+
+	res = sym_cuGetProcAddress("cuFlushGPUDirectRDMAWrites",
+			(void **) (&pfn_cuFlushGPUDirectRDMAWrites), cuda_driver_version, 0);
+	if (res != 0) {
+		rte_cuda_log(ERR, "Retrieve cuFlushGPUDirectRDMAWrites failed with %d", res);
+		return -1;
+	}
+
+	return 0;
+}
+
+/* Generate a key from a memory pointer */
+static cuda_ptr_key
+get_hash_from_ptr(void *ptr)
+{
+	return (uintptr_t) ptr;
+}
+
+static uint32_t
+mem_list_count_item(void)
+{
+	return mem_alloc_list_last_elem;
+}
+
+/* Initiate list of memory allocations if not done yet */
+static struct mem_entry *
+mem_list_add_item(void)
+{
+	/* Initiate list of memory allocations if not done yet */
+	if (mem_alloc_list_head == NULL) {
+		mem_alloc_list_head = rte_zmalloc(NULL,
+				sizeof(struct mem_entry),
+				RTE_CACHE_LINE_SIZE);
+		if (mem_alloc_list_head == NULL) {
+			rte_cuda_log(ERR, "Failed to allocate memory for memory list");
+			return NULL;
+		}
+
+		mem_alloc_list_head->next = NULL;
+		mem_alloc_list_head->prev = NULL;
+		mem_alloc_list_tail = mem_alloc_list_head;
+	} else {
+		struct mem_entry *mem_alloc_list_cur = rte_zmalloc(NULL,
+				sizeof(struct mem_entry),
+				RTE_CACHE_LINE_SIZE);
+
+		if (mem_alloc_list_cur == NULL) {
+			rte_cuda_log(ERR, "Failed to allocate memory for memory list");
+			return NULL;
+		}
+
+		mem_alloc_list_tail->next = mem_alloc_list_cur;
+		mem_alloc_list_cur->prev = mem_alloc_list_tail;
+		mem_alloc_list_tail = mem_alloc_list_tail->next;
+		mem_alloc_list_tail->next = NULL;
+	}
+
+	mem_alloc_list_last_elem++;
+
+	return mem_alloc_list_tail;
+}
+
+static struct mem_entry *
+mem_list_find_item(cuda_ptr_key pk)
+{
+	struct mem_entry *mem_alloc_list_cur = NULL;
+
+	if (mem_alloc_list_head == NULL) {
+		rte_cuda_log(ERR, "Memory list doesn't exist");
+		return NULL;
+	}
+
+	if (mem_list_count_item() == 0) {
+		rte_cuda_log(ERR, "No items in memory list");
+		return NULL;
+	}
+
+	mem_alloc_list_cur = mem_alloc_list_head;
+
+	while (mem_alloc_list_cur != NULL) {
+		if (mem_alloc_list_cur->pkey == pk)
+			return mem_alloc_list_cur;
+		mem_alloc_list_cur = mem_alloc_list_cur->next;
+	}
+
+	return mem_alloc_list_cur;
+}
+
+static int
+mem_list_del_item(cuda_ptr_key pk)
+{
+	struct mem_entry *mem_alloc_list_cur = NULL;
+
+	mem_alloc_list_cur = mem_list_find_item(pk);
+	if (mem_alloc_list_cur == NULL)
+		return -EINVAL;
+
+	/* if key is in head */
+	if (mem_alloc_list_cur->prev == NULL)
+		mem_alloc_list_head = mem_alloc_list_cur->next;
+	else {
+		mem_alloc_list_cur->prev->next = mem_alloc_list_cur->next;
+		if (mem_alloc_list_cur->next != NULL)
+			mem_alloc_list_cur->next->prev = mem_alloc_list_cur->prev;
+	}
+
+	rte_free(mem_alloc_list_cur);
+
+	mem_alloc_list_last_elem--;
+
+	return 0;
+}
+
+static int
+cuda_dev_info_get(struct rte_gpu *dev, struct rte_gpu_info *info)
+{
+	int ret = 0;
+	CUresult res;
+	struct rte_gpu_info parent_info;
+	CUexecAffinityParam affinityPrm;
+	const char *err_string;
+	struct cuda_info *private;
+	CUcontext current_ctx;
+	CUcontext input_ctx;
+
+	if (dev == NULL)
+		return -ENODEV;
+
+	/* Child initialization time probably called by rte_gpu_add_child() */
+	if (dev->mpshared->info.parent != RTE_GPU_ID_NONE && dev->mpshared->dev_private == NULL) {
+		/* Store current ctx */
+		res = pfn_cuCtxGetCurrent(&current_ctx);
+		if (res != 0) {
+			pfn_cuGetErrorString(res, &(err_string));
+			rte_cuda_log(ERR, "cuCtxGetCurrent failed with %s", err_string);
+			return -EPERM;
+		}
+
+		/* Set child ctx as current ctx */
+		input_ctx = (CUcontext)((uintptr_t)dev->mpshared->info.context);
+		res = pfn_cuCtxSetCurrent(input_ctx);
+		if (res != 0) {
+			pfn_cuGetErrorString(res, &(err_string));
+			rte_cuda_log(ERR, "cuCtxSetCurrent input failed with %s",
+					err_string);
+			return -EPERM;
+		}
+
+		/*
+		 * Ctx capacity info
+		 */
+
+		/* MPS compatible */
+		res = pfn_cuCtxGetExecAffinity(&affinityPrm, CU_EXEC_AFFINITY_TYPE_SM_COUNT);
+		if (res != 0) {
+			pfn_cuGetErrorString(res, &(err_string));
+			rte_cuda_log(ERR, "cuCtxGetExecAffinity failed with %s", err_string);
+		}
+		dev->mpshared->info.processor_count = (uint32_t)affinityPrm.param.smCount.val;
+
+		ret = rte_gpu_info_get(dev->mpshared->info.parent, &parent_info);
+		if (ret)
+			return -ENODEV;
+		dev->mpshared->info.total_memory = parent_info.total_memory;
+
+		/*
+		 * GPU Device private info
+		 */
+		dev->mpshared->dev_private = rte_zmalloc(NULL,
+							sizeof(struct cuda_info),
+							RTE_CACHE_LINE_SIZE);
+		if (dev->mpshared->dev_private == NULL) {
+			rte_cuda_log(ERR, "Failed to allocate memory for GPU process private");
+			return -EPERM;
+		}
+
+		private = (struct cuda_info *)dev->mpshared->dev_private;
+
+		res = pfn_cuCtxGetDevice(&(private->cu_dev));
+		if (res != 0) {
+			pfn_cuGetErrorString(res, &(err_string));
+			rte_cuda_log(ERR, "cuCtxGetDevice failed with %s", err_string);
+			return -EPERM;
+		}
+
+		res = pfn_cuDeviceGetName(private->gpu_name, RTE_DEV_NAME_MAX_LEN, private->cu_dev);
+		if (res != 0) {
+			pfn_cuGetErrorString(res, &(err_string));
+			rte_cuda_log(ERR, "cuDeviceGetName failed with %s", err_string);
+			return -EPERM;
+		}
+
+		/* Restore original ctx as current ctx */
+		res = pfn_cuCtxSetCurrent(current_ctx);
+		if (res != 0) {
+			pfn_cuGetErrorString(res, &(err_string));
+			rte_cuda_log(ERR, "cuCtxSetCurrent current failed with %s",
+					err_string);
+			return -EPERM;
+		}
+	}
+
+	*info = dev->mpshared->info;
+
+	return 0;
+}
+
+/*
+ * GPU Memory
+ */
+
+static int
+cuda_mem_alloc(struct rte_gpu *dev, size_t size, void **ptr)
+{
+	CUresult res;
+	const char *err_string;
+	CUcontext current_ctx;
+	CUcontext input_ctx;
+	unsigned int flag = 1;
+
+	if (dev == NULL)
+		return -ENODEV;
+	if (size == 0)
+		return -EINVAL;
+
+	/* Store current ctx */
+	res = pfn_cuCtxGetCurrent(&current_ctx);
+	if (res != 0) {
+		pfn_cuGetErrorString(res, &(err_string));
+		rte_cuda_log(ERR, "cuCtxGetCurrent failed with %s",
+				err_string);
+		return -EPERM;
+	}
+
+	/* Set child ctx as current ctx */
+	input_ctx = (CUcontext)((uintptr_t)dev->mpshared->info.context);
+	res = pfn_cuCtxSetCurrent(input_ctx);
+	if (res != 0) {
+		pfn_cuGetErrorString(res, &(err_string));
+		rte_cuda_log(ERR, "cuCtxSetCurrent input failed with %s",
+				err_string);
+		return -EPERM;
+	}
+
+	/* Get next memory list item */
+	mem_alloc_list_tail = mem_list_add_item();
+	if (mem_alloc_list_tail == NULL)
+		return -ENOMEM;
+
+	/* Allocate memory */
+	mem_alloc_list_tail->size = size;
+	res = pfn_cuMemAlloc(&(mem_alloc_list_tail->ptr_d), mem_alloc_list_tail->size);
+	if (res != 0) {
+		pfn_cuGetErrorString(res, &(err_string));
+		rte_cuda_log(ERR, "cuCtxSetCurrent current failed with %s",
+				err_string);
+		return -EPERM;
+	}
+
+	/* GPUDirect RDMA attribute required */
+	res = pfn_cuPointerSetAttribute(&flag,
+					CU_POINTER_ATTRIBUTE_SYNC_MEMOPS,
+					mem_alloc_list_tail->ptr_d);
+	if (res != 0) {
+		rte_cuda_log(ERR, "Could not set SYNC MEMOP attribute for "
+				"GPU memory at  %"PRIu32", err %d",
+				(uint32_t) mem_alloc_list_tail->ptr_d, res);
+		return -EPERM;
+	}
+
+	mem_alloc_list_tail->pkey = get_hash_from_ptr((void *) mem_alloc_list_tail->ptr_d);
+	mem_alloc_list_tail->ptr_h = NULL;
+	mem_alloc_list_tail->size = size;
+	mem_alloc_list_tail->dev = dev;
+	mem_alloc_list_tail->ctx = (CUcontext)((uintptr_t)dev->mpshared->info.context);
+	mem_alloc_list_tail->mtype = GPU_MEM;
+
+	/* Restore original ctx as current ctx */
+	res = pfn_cuCtxSetCurrent(current_ctx);
+	if (res != 0) {
+		pfn_cuGetErrorString(res, &(err_string));
+		rte_cuda_log(ERR, "cuCtxSetCurrent current failed with %s",
+				err_string);
+		return -EPERM;
+	}
+
+	*ptr = (void *) mem_alloc_list_tail->ptr_d;
+
+	return 0;
+}
+
+static int
+cuda_mem_register(struct rte_gpu *dev, size_t size, void *ptr)
+{
+	CUresult res;
+	const char *err_string;
+	CUcontext current_ctx;
+	CUcontext input_ctx;
+	unsigned int flag = 1;
+	int use_ptr_h = 0;
+
+	if (dev == NULL)
+		return -ENODEV;
+
+	if (size == 0 || ptr == NULL)
+		return -EINVAL;
+
+	/* Store current ctx */
+	res = pfn_cuCtxGetCurrent(&current_ctx);
+	if (res != 0) {
+		pfn_cuGetErrorString(res, &(err_string));
+		rte_cuda_log(ERR, "cuCtxGetCurrent failed with %s",
+				err_string);
+		return -EPERM;
+	}
+
+	/* Set child ctx as current ctx */
+	input_ctx = (CUcontext)((uintptr_t)dev->mpshared->info.context);
+	res = pfn_cuCtxSetCurrent(input_ctx);
+	if (res != 0) {
+		pfn_cuGetErrorString(res, &(err_string));
+		rte_cuda_log(ERR, "cuCtxSetCurrent input failed with %s",
+				err_string);
+		return -EPERM;
+	}
+
+	/* Get next memory list item */
+	mem_alloc_list_tail = mem_list_add_item();
+	if (mem_alloc_list_tail == NULL)
+		return -ENOMEM;
+
+	/* Allocate memory */
+	mem_alloc_list_tail->size = size;
+	mem_alloc_list_tail->ptr_h = ptr;
+
+	res = pfn_cuMemHostRegister(mem_alloc_list_tail->ptr_h,
+				mem_alloc_list_tail->size,
+				CU_MEMHOSTREGISTER_PORTABLE | CU_MEMHOSTREGISTER_DEVICEMAP);
+	if (res != 0) {
+		pfn_cuGetErrorString(res, &(err_string));
+		rte_cuda_log(ERR, "cuMemHostRegister failed with %s ptr %p size %zd",
+				err_string,
+				mem_alloc_list_tail->ptr_h,
+				mem_alloc_list_tail->size);
+		return -EPERM;
+	}
+
+	res = pfn_cuDeviceGetAttribute(&(use_ptr_h),
+			CU_DEVICE_ATTRIBUTE_CAN_USE_HOST_POINTER_FOR_REGISTERED_MEM,
+			((struct cuda_info *)(dev->mpshared->dev_private))->cu_dev);
+	if (res != 0) {
+		pfn_cuGetErrorString(res, &(err_string));
+		rte_cuda_log(ERR, "cuDeviceGetAttribute failed with %s",
+				err_string);
+		return -EPERM;
+	}
+
+	if (use_ptr_h == 0) {
+		res = pfn_cuMemHostGetDevicePointer(&(mem_alloc_list_tail->ptr_d),
+				mem_alloc_list_tail->ptr_h, 0);
+		if (res != 0) {
+			pfn_cuGetErrorString(res, &(err_string));
+			rte_cuda_log(ERR, "cuMemHostGetDevicePointer failed with %s",
+					err_string);
+			return -EPERM;
+		}
+
+		if ((uintptr_t) mem_alloc_list_tail->ptr_d != (uintptr_t) mem_alloc_list_tail->ptr_h) {
+			rte_cuda_log(ERR, "Host input pointer is different wrt GPU registered pointer");
+			return -ENOTSUP;
+		}
+	} else {
+		mem_alloc_list_tail->ptr_d = (CUdeviceptr) mem_alloc_list_tail->ptr_h;
+	}
+
+	/* GPUDirect RDMA attribute required */
+	res = pfn_cuPointerSetAttribute(&flag,
+					CU_POINTER_ATTRIBUTE_SYNC_MEMOPS,
+					mem_alloc_list_tail->ptr_d);
+	if (res != 0) {
+		rte_cuda_log(ERR, "Could not set SYNC MEMOP attribute for GPU memory at %"PRIu32", err %d",
+				(uint32_t) mem_alloc_list_tail->ptr_d, res);
+		return -EPERM;
+	}
+
+	mem_alloc_list_tail->pkey = get_hash_from_ptr((void *) mem_alloc_list_tail->ptr_h);
+	mem_alloc_list_tail->size = size;
+	mem_alloc_list_tail->dev = dev;
+	mem_alloc_list_tail->ctx = (CUcontext)((uintptr_t)dev->mpshared->info.context);
+	mem_alloc_list_tail->mtype = CPU_REGISTERED;
+
+	/* Restore original ctx as current ctx */
+	res = pfn_cuCtxSetCurrent(current_ctx);
+	if (res != 0) {
+		pfn_cuGetErrorString(res, &(err_string));
+		rte_cuda_log(ERR, "cuCtxSetCurrent current failed with %s",
+				err_string);
+		return -EPERM;
+	}
+
+	return 0;
+}
+
+static int
+cuda_mem_free(struct rte_gpu *dev, void *ptr)
+{
+	CUresult res;
+	struct mem_entry *mem_item;
+	const char *err_string;
+	cuda_ptr_key hk;
+
+	if (dev == NULL)
+		return -ENODEV;
+
+	if (ptr == NULL)
+		return -EINVAL;
+
+	hk = get_hash_from_ptr((void *) ptr);
+
+	mem_item = mem_list_find_item(hk);
+	if (mem_item == NULL) {
+		rte_cuda_log(ERR, "Memory address 0x%p not found in driver memory", ptr);
+		return -EPERM;
+	}
+
+	if (mem_item->mtype == GPU_MEM) {
+		res = pfn_cuMemFree(mem_item->ptr_d);
+		if (res != 0) {
+			pfn_cuGetErrorString(res, &(err_string));
+			rte_cuda_log(ERR, "cuMemFree current failed with %s",
+					err_string);
+			return -EPERM;
+		}
+
+		return mem_list_del_item(hk);
+	}
+
+	rte_cuda_log(ERR, "Memory type %d not supported", mem_item->mtype);
+
+	return -EPERM;
+}
+
+static int
+cuda_mem_unregister(struct rte_gpu *dev, void *ptr)
+{
+	CUresult res;
+	struct mem_entry *mem_item;
+	const char *err_string;
+	cuda_ptr_key hk;
+
+	if (dev == NULL)
+		return -ENODEV;
+
+	if (ptr == NULL)
+		return -EINVAL;
+
+	hk = get_hash_from_ptr((void *) ptr);
+
+	mem_item = mem_list_find_item(hk);
+	if (mem_item == NULL) {
+		rte_cuda_log(ERR, "Memory address 0x%p not found in driver memory", ptr);
+		return -EPERM;
+	}
+
+	if (mem_item->mtype == CPU_REGISTERED) {
+		res = pfn_cuMemHostUnregister(ptr);
+		if (res != 0) {
+			pfn_cuGetErrorString(res, &(err_string));
+			rte_cuda_log(ERR, "cuMemHostUnregister current failed with %s",
+					err_string);
+			return -EPERM;
+		}
+
+		return mem_list_del_item(hk);
+	}
+
+	rte_cuda_log(ERR, "Memory type %d not supported", mem_item->mtype);
+
+	return -EPERM;
+}
+
+static int
+cuda_dev_close(struct rte_gpu *dev)
+{
+	if (dev == NULL)
+		return -EINVAL;
+
+	rte_free(dev->mpshared->dev_private);
+
+	return 0;
+}
+
+static int
+cuda_wmb(struct rte_gpu *dev)
+{
+	CUresult res;
+	const char *err_string;
+	CUcontext current_ctx;
+	CUcontext input_ctx;
+	struct cuda_info *private;
+
+	if (dev == NULL)
+		return -ENODEV;
+
+	private = (struct cuda_info *)dev->mpshared->dev_private;
+
+	if (private->gdr_write_ordering != CU_GPU_DIRECT_RDMA_WRITES_ORDERING_NONE) {
+		/*
+		 * No need to explicitly force the write ordering because
+		 * the device natively supports it
+		 */
+		return 0;
+	}
+
+	if (private->gdr_flush_type != CU_FLUSH_GPU_DIRECT_RDMA_WRITES_OPTION_HOST) {
+		/*
+		 * Can't flush GDR writes with cuFlushGPUDirectRDMAWrites CUDA function.
+		 * Application needs to use alternative methods.
+		 */
+		rte_cuda_log(WARNING, "Can't flush GDR writes with cuFlushGPUDirectRDMAWrites CUDA function."
+				"Application needs to use alternative methods.");
+		return -ENOTSUP;
+	}
+
+	/* Store current ctx */
+	res = pfn_cuCtxGetCurrent(&current_ctx);
+	if (res != 0) {
+		pfn_cuGetErrorString(res, &(err_string));
+		rte_cuda_log(ERR, "cuCtxGetCurrent failed with %s",
+				err_string);
+		return -EPERM;
+	}
+
+	/* Set child ctx as current ctx */
+	input_ctx = (CUcontext)((uintptr_t)dev->mpshared->info.context);
+	res = pfn_cuCtxSetCurrent(input_ctx);
+	if (res != 0) {
+		pfn_cuGetErrorString(res, &(err_string));
+		rte_cuda_log(ERR, "cuCtxSetCurrent input failed with %s",
+				err_string);
+		return -EPERM;
+	}
+
+	res = pfn_cuFlushGPUDirectRDMAWrites(CU_FLUSH_GPU_DIRECT_RDMA_WRITES_TARGET_CURRENT_CTX,
+					CU_FLUSH_GPU_DIRECT_RDMA_WRITES_TO_ALL_DEVICES);
+	if (res != 0) {
+		pfn_cuGetErrorString(res, &(err_string));
+		rte_cuda_log(ERR, "cuFlushGPUDirectRDMAWrites current failed with %s",
+				err_string);
+		return -EPERM;
+	}
+
+	/* Restore original ctx as current ctx */
+	res = pfn_cuCtxSetCurrent(current_ctx);
+	if (res != 0) {
+		pfn_cuGetErrorString(res, &(err_string));
+		rte_cuda_log(ERR, "cuCtxSetCurrent current failed with %s",
+				err_string);
+		return -EPERM;
+	}
+
+	return 0;
+}
+
+static int
+cuda_gpu_probe(__rte_unused struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
+{
+	struct rte_gpu *dev = NULL;
+	CUresult res;
+	CUdevice cu_dev_id;
+	CUcontext pctx;
+	char dev_name[RTE_DEV_NAME_MAX_LEN];
+	const char *err_string;
+	int processor_count = 0;
+	struct cuda_info *private;
+
+	if (pci_dev == NULL) {
+		rte_cuda_log(ERR, "NULL PCI device");
+		return -EINVAL;
+	}
+
+	rte_pci_device_name(&pci_dev->addr, dev_name, sizeof(dev_name));
+
+	/* Allocate memory to be used privately by drivers */
+	dev = rte_gpu_allocate(pci_dev->device.name);
+	if (dev == NULL)
+		return -ENODEV;
+
+	/* Initialize values only for the first CUDA driver call */
+	if (dev->mpshared->info.dev_id == 0) {
+		mem_alloc_list_head = NULL;
+		mem_alloc_list_tail = NULL;
+		mem_alloc_list_last_elem = 0;
+
+		/* Load libcuda.so library */
+		if (cuda_loader()) {
+			rte_cuda_log(ERR, "CUDA Driver library not found");
+			return -ENOTSUP;
+		}
+
+		/* Load initial CUDA functions */
+		if (cuda_sym_func_loader()) {
+			rte_cuda_log(ERR, "CUDA functions not found in library");
+			return -ENOTSUP;
+		}
+
+		/*
+		 * Required to initialize the CUDA Driver.
+		 * Multiple calls of cuInit() will return immediately
+		 * without making any relevant change
+		 */
+		sym_cuInit(0);
+
+		res = sym_cuDriverGetVersion(&cuda_driver_version);
+		if (res != 0) {
+			rte_cuda_log(ERR, "cuDriverGetVersion failed with %d", res);
+			return -ENOTSUP;
+		}
+
+		if (cuda_driver_version < CUDA_DRIVER_MIN_VERSION) {
+			rte_cuda_log(ERR, "CUDA Driver version found is %d. "
+					"Minimum requirement is %d",
+					cuda_driver_version, CUDA_DRIVER_MIN_VERSION);
+			return -ENOTSUP;
+		}
+
+		if (cuda_pfn_func_loader()) {
+			rte_cuda_log(ERR, "CUDA PFN functions not found in library");
+			return -ENOTSUP;
+		}
+	}
+
+	/* Fill HW specific part of device structure */
+	dev->device = &pci_dev->device;
+	dev->mpshared->info.numa_node = pci_dev->device.numa_node;
+
+	/* Get NVIDIA GPU Device descriptor */
+	res = pfn_cuDeviceGetByPCIBusId(&cu_dev_id, dev->device->name);
+	if (res != 0) {
+		pfn_cuGetErrorString(res, &(err_string));
+		rte_cuda_log(ERR, "cuDeviceGetByPCIBusId name %s failed with %d: %s",
+				dev->device->name, res, err_string);
+		return -EPERM;
+	}
+
+	res = pfn_cuDevicePrimaryCtxRetain(&pctx, cu_dev_id);
+	if (res != 0) {
+		pfn_cuGetErrorString(res, &(err_string));
+		rte_cuda_log(ERR, "cuDevicePrimaryCtxRetain name %s failed with %d: %s",
+				dev->device->name, res, err_string);
+		return -EPERM;
+	}
+
+	res = pfn_cuCtxGetApiVersion(pctx, &cuda_api_version);
+	if (res != 0) {
+		rte_cuda_log(ERR, "cuCtxGetApiVersion failed with %d", res);
+		return -ENOTSUP;
+	}
+
+	if (cuda_api_version < CUDA_API_MIN_VERSION) {
+		rte_cuda_log(ERR, "CUDA API version found is %d Minimum requirement is %d",
+				cuda_api_version, CUDA_API_MIN_VERSION);
+		return -ENOTSUP;
+	}
+
+	dev->mpshared->info.context = (uint64_t) pctx;
+
+	/*
+	 * GPU Device generic info
+	 */
+
+	/* Processor count */
+	res = pfn_cuDeviceGetAttribute(&(processor_count),
+			CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT,
+			cu_dev_id);
+	if (res != 0) {
+		pfn_cuGetErrorString(res, &(err_string));
+		rte_cuda_log(ERR, "cuDeviceGetAttribute failed with %s",
+				err_string);
+		return -EPERM;
+	}
+	dev->mpshared->info.processor_count = (uint32_t)processor_count;
+
+	/* Total memory */
+	res = pfn_cuDeviceTotalMem(&dev->mpshared->info.total_memory, cu_dev_id);
+	if (res != 0) {
+		pfn_cuGetErrorString(res, &(err_string));
+		rte_cuda_log(ERR, "cuDeviceTotalMem failed with %s",
+				err_string);
+		return -EPERM;
+	}
+
+	/*
+	 * GPU Device private info
+	 */
+	dev->mpshared->dev_private = rte_zmalloc(NULL,
+			sizeof(struct cuda_info),
+			RTE_CACHE_LINE_SIZE);
+	if (dev->mpshared->dev_private == NULL) {
+		rte_cuda_log(ERR, "Failed to allocate memory for GPU process private");
+		return -ENOMEM;
+	}
+
+	private = (struct cuda_info *)dev->mpshared->dev_private;
+	private->cu_dev = cu_dev_id;
+	res = pfn_cuDeviceGetName(private->gpu_name,
+			RTE_DEV_NAME_MAX_LEN,
+			cu_dev_id);
+	if (res != 0) {
+		pfn_cuGetErrorString(res, &(err_string));
+		rte_cuda_log(ERR, "cuDeviceGetName failed with %s",
+				err_string);
+		return -EPERM;
+	}
+
+	res = pfn_cuDeviceGetAttribute(&(private->gdr_supported),
+			CU_DEVICE_ATTRIBUTE_GPU_DIRECT_RDMA_SUPPORTED,
+			cu_dev_id);
+	if (res != 0) {
+		pfn_cuGetErrorString(res, &(err_string));
+		rte_cuda_log(ERR, "cuDeviceGetAttribute failed with %s",
+				err_string);
+		return -EPERM;
+	}
+
+	if (private->gdr_supported == 0)
+		rte_cuda_log(WARNING, "GPU %s doesn't support GPUDirect RDMA",
+				pci_dev->device.name);
+
+	res = pfn_cuDeviceGetAttribute(&(private->gdr_write_ordering),
+			CU_DEVICE_ATTRIBUTE_GPU_DIRECT_RDMA_WRITES_ORDERING,
+			cu_dev_id);
+	if (res != 0) {
+		pfn_cuGetErrorString(res, &(err_string));
+		rte_cuda_log(ERR,
+				"cuDeviceGetAttribute failed with %s",
+				err_string);
+		return -EPERM;
+	}
+
+	if (private->gdr_write_ordering == CU_GPU_DIRECT_RDMA_WRITES_ORDERING_NONE) {
+		res = pfn_cuDeviceGetAttribute(&(private->gdr_flush_type),
+					CU_DEVICE_ATTRIBUTE_GPU_DIRECT_RDMA_FLUSH_WRITES_OPTIONS,
+					cu_dev_id);
+		if (res != 0) {
+			pfn_cuGetErrorString(res, &(err_string));
+			rte_cuda_log(ERR, "cuDeviceGetAttribute failed with %s",
+					err_string);
+			return -EPERM;
+		}
+
+		if (private->gdr_flush_type != CU_FLUSH_GPU_DIRECT_RDMA_WRITES_OPTION_HOST)
+			rte_cuda_log(ERR, "GPUDirect RDMA flush writes API is not supported");
+	}
+
+	dev->ops.dev_info_get = cuda_dev_info_get;
+	dev->ops.dev_close = cuda_dev_close;
+	dev->ops.mem_alloc = cuda_mem_alloc;
+	dev->ops.mem_free = cuda_mem_free;
+	dev->ops.mem_register = cuda_mem_register;
+	dev->ops.mem_unregister = cuda_mem_unregister;
+	dev->ops.wmb = cuda_wmb;
+
+	rte_gpu_complete_new(dev);
+
+	rte_cuda_debug("dev id = %u name = %s",
+			dev->mpshared->info.dev_id, private->gpu_name);
+
+	return 0;
+}
+
+static int
+cuda_gpu_remove(struct rte_pci_device *pci_dev)
+{
+	struct rte_gpu *dev;
+	int ret;
+	uint8_t gpu_id;
+
+	if (pci_dev == NULL)
+		return -EINVAL;
+
+	dev = rte_gpu_get_by_name(pci_dev->device.name);
+	if (dev == NULL) {
+		rte_cuda_log(ERR, "Couldn't find HW dev \"%s\" to uninitialise it",
+				pci_dev->device.name);
+		return -ENODEV;
+	}
+	gpu_id = dev->mpshared->info.dev_id;
+
+	/* release dev from library */
+	ret = rte_gpu_release(dev);
+	if (ret)
+		rte_cuda_log(ERR, "Device %i failed to uninit: %i", gpu_id, ret);
+
+	rte_cuda_debug("Destroyed dev = %u", gpu_id);
+
+	return 0;
+}
+
+static struct rte_pci_driver rte_cuda_driver = {
+	.id_table = pci_id_cuda_map,
+	.drv_flags = RTE_PCI_DRV_WC_ACTIVATE,
+	.probe = cuda_gpu_probe,
+	.remove = cuda_gpu_remove,
+};
+
+RTE_PMD_REGISTER_PCI(gpu_cuda, rte_cuda_driver);
+RTE_PMD_REGISTER_PCI_TABLE(gpu_cuda, pci_id_cuda_map);
+RTE_PMD_REGISTER_KMOD_DEP(gpu_cuda, "* nvidia & (nv_peer_mem | nvpeer_mem)");
diff --git a/drivers/gpu/cuda/meson.build b/drivers/gpu/cuda/meson.build
new file mode 100644
index 0000000000..4bb3e14cd2
--- /dev/null
+++ b/drivers/gpu/cuda/meson.build
@@ -0,0 +1,18 @@
+# SPDX-License-Identifier: BSD-3-Clause
+# Copyright (c) 2021 NVIDIA Corporation & Affiliates
+
+if not is_linux
+        build = false
+        reason = 'only supported on Linux'
+endif
+
+cuda_h_found = cc.has_header('cuda.h')
+cuda_typeh_found = cc.has_header('cudaTypedefs.h')
+
+if not cuda_h_found or not cuda_typeh_found
+        build = false
+        reason = 'missing dependency cuda headers cuda.h and cudaTypedefs.h'
+endif
+
+deps += ['gpudev','pci','bus_pci']
+sources = files('cuda.c')
diff --git a/drivers/gpu/cuda/version.map b/drivers/gpu/cuda/version.map
new file mode 100644
index 0000000000..4a76d1d52d
--- /dev/null
+++ b/drivers/gpu/cuda/version.map
@@ -0,0 +1,3 @@
+DPDK_21 {
+	local: *;
+};
diff --git a/drivers/gpu/meson.build b/drivers/gpu/meson.build
index e51ad3381b..601bedcd61 100644
--- a/drivers/gpu/meson.build
+++ b/drivers/gpu/meson.build
@@ -1,4 +1,4 @@
 # SPDX-License-Identifier: BSD-3-Clause
 # Copyright (c) 2021 NVIDIA Corporation & Affiliates
 
-drivers = []
+drivers = [ 'cuda' ]
-- 
2.17.1