[dpdk-dev] [PATCH v3 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.
Funcitonalities 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     | 110 +++++
 doc/guides/gpus/index.rst    |   1 +
 drivers/gpu/cuda/cuda.c      | 801 +++++++++++++++++++++++++++++++++++
 drivers/gpu/cuda/meson.build |  13 +
 drivers/gpu/cuda/version.map |   3 +
 drivers/gpu/meson.build      |   2 +-
 6 files changed, 929 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..64a78bf1e1
--- /dev/null
+++ b/doc/guides/gpus/cuda.rst
@@ -0,0 +1,110 @@
+.. 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>`__.
+
+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, a 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:
+
+- 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 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 it's 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/drivers/gpu/cuda/cuda.c b/drivers/gpu/cuda/cuda.c
new file mode 100644
index 0000000000..d3d57492db
--- /dev/null
+++ b/drivers/gpu/cuda/cuda.c
@@ -0,0 +1,801 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright (c) 2021 NVIDIA Corporation & Affiliates
+ */
+
+#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>
+
+/* 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 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_gpu_cuda_log(level, fmt, ...) \
+	rte_log(RTE_LOG_ ## level, cuda_logtype, fmt "\n", ##__VA_ARGS__)
+
+#define rte_gpu_cuda_log_debug(fmt, ...) \
+	rte_gpu_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_V100_32GB_DEVICE_ID)
+	},
+	/* {.device_id = 0}, ?? */
+};
+
+/* Device private info */
+struct cuda_info {
+	char gpu_name[RTE_DEV_NAME_MAX_LEN];
+	CUdevice cu_dev;
+};
+
+/* 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;
+
+/* 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_gpu_cuda_log(ERR, "Failed to allocate memory for memory list.\n");
+			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_gpu_cuda_log(ERR, "Failed to allocate memory for memory list.\n");
+			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_gpu_cuda_log(ERR, "Memory list doesn't exist\n");
+		return NULL;
+	}
+
+	if (mem_list_count_item() == 0) {
+		rte_gpu_cuda_log(ERR, "No items in memory list\n");
+		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 -EINVAL;
+
+	/* 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 = cuCtxGetCurrent(&current_ctx);
+		if (res != CUDA_SUCCESS) {
+			cuGetErrorString(res, &(err_string));
+			rte_gpu_cuda_log(ERR, "cuCtxGetCurrent failed with %s.\n", err_string);
+
+			return -1;
+		}
+
+		/* Set child ctx as current ctx */
+		input_ctx = (CUcontext)dev->mpshared->info.context;
+		res = cuCtxSetCurrent(input_ctx);
+		if (res != CUDA_SUCCESS) {
+			cuGetErrorString(res, &(err_string));
+			rte_gpu_cuda_log(ERR, "cuCtxSetCurrent input failed with %s.\n", err_string);
+
+			return -1;
+		}
+
+		/*
+		 * Ctx capacity info
+		 */
+
+		/* MPS compatible */
+		res = cuCtxGetExecAffinity(&affinityPrm, CU_EXEC_AFFINITY_TYPE_SM_COUNT);
+		if (res != CUDA_SUCCESS) {
+			cuGetErrorString(res, &(err_string));
+			rte_gpu_cuda_log(ERR, "cuCtxGetExecAffinity failed with %s.\n", 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_gpu_cuda_log(ERR, "Failed to allocate memory for GPU process private.\n");
+
+			return -1;
+		}
+
+		private = (struct cuda_info *)dev->mpshared->dev_private;
+
+		res = cuCtxGetDevice(&(private->cu_dev));
+		if (res != CUDA_SUCCESS) {
+			cuGetErrorString(res, &(err_string));
+			rte_gpu_cuda_log(ERR, "cuCtxGetDevice failed with %s.\n", err_string);
+
+			return -1;
+		}
+
+		res = cuDeviceGetName(private->gpu_name, RTE_DEV_NAME_MAX_LEN, private->cu_dev);
+		if (res != CUDA_SUCCESS) {
+			cuGetErrorString(res, &(err_string));
+			rte_gpu_cuda_log(ERR, "cuDeviceGetName failed with %s.\n", err_string);
+
+			return -1;
+		}
+
+		/* Restore original ctx as current ctx */
+		res = cuCtxSetCurrent(current_ctx);
+		if (res != CUDA_SUCCESS) {
+			cuGetErrorString(res, &(err_string));
+			rte_gpu_cuda_log(ERR, "cuCtxSetCurrent current failed with %s.\n", err_string);
+
+			return -1;
+		}
+	}
+
+	*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 || size == 0)
+		return -EINVAL;
+
+	/* Store current ctx */
+	res = cuCtxGetCurrent(&current_ctx);
+	if (res != CUDA_SUCCESS) {
+		cuGetErrorString(res, &(err_string));
+		rte_gpu_cuda_log(ERR, "cuCtxGetCurrent failed with %s.\n", err_string);
+
+		return -1;
+	}
+
+	/* Set child ctx as current ctx */
+	input_ctx = (CUcontext)dev->mpshared->info.context;
+	res = cuCtxSetCurrent(input_ctx);
+	if (res != CUDA_SUCCESS) {
+		cuGetErrorString(res, &(err_string));
+		rte_gpu_cuda_log(ERR, "cuCtxSetCurrent input failed with %s.\n", err_string);
+
+		return -1;
+	}
+
+	/* 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 = cuMemAlloc(&(mem_alloc_list_tail->ptr_d), mem_alloc_list_tail->size);
+	if (res != CUDA_SUCCESS) {
+		cuGetErrorString(res, &(err_string));
+		rte_gpu_cuda_log(ERR,
+				"cuCtxSetCurrent current failed with %s.\n",
+				err_string);
+
+		return -1;
+	}
+
+	/* GPUDirect RDMA attribute required */
+	res = cuPointerSetAttribute(&flag,
+					CU_POINTER_ATTRIBUTE_SYNC_MEMOPS,
+					mem_alloc_list_tail->ptr_d);
+	if (res != CUDA_SUCCESS) {
+		rte_gpu_cuda_log(ERR,
+				"Could not set SYNC MEMOP attribute for GPU memory at %llx , err %d\n",
+				mem_alloc_list_tail->ptr_d, res);
+		return -1;
+	}
+
+	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)dev->mpshared->info.context;
+	mem_alloc_list_tail->mtype = GPU_MEM;
+
+	/* Restore original ctx as current ctx */
+	res = cuCtxSetCurrent(current_ctx);
+	if (res != CUDA_SUCCESS) {
+		cuGetErrorString(res, &(err_string));
+		rte_gpu_cuda_log(ERR, "cuCtxSetCurrent current failed with %s.\n", err_string);
+
+		return -1;
+	}
+
+	*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 || size == 0 || ptr == NULL)
+		return -EINVAL;
+
+	/* Store current ctx */
+	res = cuCtxGetCurrent(&current_ctx);
+	if (res != CUDA_SUCCESS) {
+		cuGetErrorString(res, &(err_string));
+		rte_gpu_cuda_log(ERR, "cuCtxGetCurrent failed with %s.\n", err_string);
+
+		return -1;
+	}
+
+	/* Set child ctx as current ctx */
+	input_ctx = (CUcontext)dev->mpshared->info.context;
+	res = cuCtxSetCurrent(input_ctx);
+	if (res != CUDA_SUCCESS) {
+		cuGetErrorString(res, &(err_string));
+		rte_gpu_cuda_log(ERR, "cuCtxSetCurrent input failed with %s.\n", err_string);
+
+		return -1;
+	}
+
+	/* 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 = cuMemHostRegister(mem_alloc_list_tail->ptr_h, mem_alloc_list_tail->size, CU_MEMHOSTREGISTER_PORTABLE | CU_MEMHOSTREGISTER_DEVICEMAP);
+	if (res != CUDA_SUCCESS) {
+		cuGetErrorString(res, &(err_string));
+		rte_gpu_cuda_log(ERR,
+				"cuMemHostRegister failed with %s ptr %p size %zd.\n",
+				err_string, mem_alloc_list_tail->ptr_h, mem_alloc_list_tail->size);
+
+		return -1;
+	}
+
+	res = 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 != CUDA_SUCCESS) {
+		cuGetErrorString(res, &(err_string));
+		rte_gpu_cuda_log(ERR, "cuDeviceGetAttribute failed with %s.\n",
+					err_string
+			);
+
+		return -1;
+	}
+
+	if (use_ptr_h == 0) {
+		res = cuMemHostGetDevicePointer(&(mem_alloc_list_tail->ptr_d),
+						mem_alloc_list_tail->ptr_h,
+						0);
+		if (res != CUDA_SUCCESS) {
+			cuGetErrorString(res, &(err_string));
+			rte_gpu_cuda_log(ERR,
+					"cuMemHostGetDevicePointer failed with %s.\n",
+					err_string);
+
+			return -1;
+		}
+
+		if ((uintptr_t) mem_alloc_list_tail->ptr_d != (uintptr_t) mem_alloc_list_tail->ptr_h) {
+			rte_gpu_cuda_log(ERR, "Host input pointer is different wrt GPU registered pointer\n");
+			return -1;
+		}
+	} else {
+		mem_alloc_list_tail->ptr_d = (CUdeviceptr) mem_alloc_list_tail->ptr_h;
+	}
+
+	/* GPUDirect RDMA attribute required */
+	res = cuPointerSetAttribute(&flag,
+					CU_POINTER_ATTRIBUTE_SYNC_MEMOPS,
+					mem_alloc_list_tail->ptr_d);
+	if (res != CUDA_SUCCESS) {
+		rte_gpu_cuda_log(ERR,
+				"Could not set SYNC MEMOP attribute for GPU memory at %llx , err %d\n",
+				mem_alloc_list_tail->ptr_d, res);
+		return -1;
+	}
+
+	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)dev->mpshared->info.context;
+	mem_alloc_list_tail->mtype = CPU_REGISTERED;
+
+	/* Restore original ctx as current ctx */
+	res = cuCtxSetCurrent(current_ctx);
+	if (res != CUDA_SUCCESS) {
+		cuGetErrorString(res, &(err_string));
+		rte_gpu_cuda_log(ERR,
+				"cuCtxSetCurrent current failed with %s.\n",
+				err_string);
+
+		return -1;
+	}
+
+	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 || ptr == NULL)
+		return -EINVAL;
+
+	hk = get_hash_from_ptr((void *) ptr);
+
+	mem_item = mem_list_find_item(hk);
+	if (mem_item == NULL) {
+		rte_gpu_cuda_log(ERR, "Memory address 0x%p not found in driver memory\n", ptr);
+		return -1;
+	}
+
+	if (mem_item->mtype == GPU_MEM) {
+		res = cuMemFree(mem_item->ptr_d);
+		if (res != CUDA_SUCCESS) {
+			cuGetErrorString(res, &(err_string));
+			rte_gpu_cuda_log(ERR, "cuMemFree current failed with %s.\n", err_string);
+
+			return -1;
+		}
+
+		return mem_list_del_item(hk);
+	}
+
+	rte_gpu_cuda_log(ERR, "Memory type %d not supported\n", mem_item->mtype);
+	return -1;
+}
+
+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 || ptr == NULL)
+		return -EINVAL;
+
+	hk = get_hash_from_ptr((void *) ptr);
+
+	mem_item = mem_list_find_item(hk);
+	if (mem_item == NULL) {
+		rte_gpu_cuda_log(ERR, "Memory address 0x%p not nd in driver memory\n", ptr);
+		return -1;
+	}
+
+	if (mem_item->mtype == CPU_REGISTERED) {
+		res = cuMemHostUnregister(ptr);
+		if (res != CUDA_SUCCESS) {
+			cuGetErrorString(res, &(err_string));
+			rte_gpu_cuda_log(ERR,
+					"cuMemHostUnregister current failed with %s.\n",
+					err_string);
+
+			return -1;
+		}
+
+		return mem_list_del_item(hk);
+	}
+
+	rte_gpu_cuda_log(ERR, "Memory type %d not supported\n", mem_item->mtype);
+	return -1;
+}
+
+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;
+
+	if (dev == NULL)
+		return -EINVAL;
+
+	/* Store current ctx */
+	res = cuCtxGetCurrent(&current_ctx);
+	if (res != CUDA_SUCCESS) {
+		cuGetErrorString(res, &(err_string));
+		rte_gpu_cuda_log(ERR, "cuCtxGetCurrent failed with %s.\n", err_string);
+
+		return -1;
+	}
+
+	/* Set child ctx as current ctx */
+	input_ctx = (CUcontext)dev->mpshared->info.context;
+	res = cuCtxSetCurrent(input_ctx);
+	if (res != CUDA_SUCCESS) {
+		cuGetErrorString(res, &(err_string));
+		rte_gpu_cuda_log(ERR, "cuCtxSetCurrent input failed with %s.\n", err_string);
+
+		return -1;
+	}
+
+	res = cuFlushGPUDirectRDMAWrites(CU_FLUSH_GPU_DIRECT_RDMA_WRITES_TARGET_CURRENT_CTX, CU_FLUSH_GPU_DIRECT_RDMA_WRITES_TO_ALL_DEVICES);
+	if (res != CUDA_SUCCESS) {
+		cuGetErrorString(res, &(err_string));
+		rte_gpu_cuda_log(ERR, "cuFlushGPUDirectRDMAWrites current failed with %s.\n", err_string);
+
+		return -1;
+	}
+
+	/* Restore original ctx as current ctx */
+	res = cuCtxSetCurrent(current_ctx);
+	if (res != CUDA_SUCCESS) {
+		cuGetErrorString(res, &(err_string));
+		rte_gpu_cuda_log(ERR, "cuCtxSetCurrent current failed with %s.\n", err_string);
+
+		return -1;
+	}
+
+	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_gpu_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;
+	}
+
+	/* Fill HW specific part of device structure */
+	dev->device = &pci_dev->device;
+	dev->mpshared->info.numa_node = pci_dev->device.numa_node;
+
+	/*
+	 * GPU Device init
+	 */
+
+	/*
+	 * Required to initialize the CUDA Driver.
+	 * Multiple calls of cuInit() will return immediately
+	 * without making any relevant change
+	 */
+	cuInit(0);
+
+	/* Get NVIDIA GPU Device descriptor */
+	res = cuDeviceGetByPCIBusId(&cu_dev_id, dev->device->name);
+	if (res != CUDA_SUCCESS) {
+		cuGetErrorString(res, &(err_string));
+		rte_gpu_cuda_log(ERR,
+				"cuDeviceGetByPCIBusId name %s failed with %d: %s.\n",
+				dev->device->name, res, err_string);
+
+		return -1;
+	}
+
+	res = cuDevicePrimaryCtxRetain(&pctx, cu_dev_id);
+	if (res != CUDA_SUCCESS) {
+		cuGetErrorString(res, &(err_string));
+		rte_gpu_cuda_log(ERR,
+				"cuDevicePrimaryCtxRetain name %s failed with %d: %s.\n",
+				dev->device->name, res, err_string);
+
+		return -1;
+	}
+
+	dev->mpshared->info.context = (uint64_t) pctx;
+
+	/*
+	 * GPU Device generic info
+	 */
+
+	/* Processor count */
+	res = cuDeviceGetAttribute(&(processor_count),
+					CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT,
+					cu_dev_id);
+	if (res != CUDA_SUCCESS) {
+		cuGetErrorString(res, &(err_string));
+		rte_gpu_cuda_log(ERR,
+				"cuDeviceGetAttribute failed with %s.\n",
+				err_string);
+
+		return -1;
+	}
+	dev->mpshared->info.processor_count = (uint32_t)processor_count;
+
+	/* Total memory */
+	res = cuDeviceTotalMem(&dev->mpshared->info.total_memory, cu_dev_id);
+	if (res != CUDA_SUCCESS) {
+		cuGetErrorString(res, &(err_string));
+		rte_gpu_cuda_log(ERR,
+				"cuDeviceTotalMem failed with %s.\n",
+				err_string);
+
+		return -1;
+	}
+
+	/*
+	 * 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_gpu_cuda_log(ERR,
+				"Failed to allocate memory for GPU process private.\n");
+
+		return -1;
+	}
+
+	private = (struct cuda_info *)dev->mpshared->dev_private;
+	private->cu_dev = cu_dev_id;
+	res = cuDeviceGetName(private->gpu_name,
+				RTE_DEV_NAME_MAX_LEN,
+				cu_dev_id);
+	if (res != CUDA_SUCCESS) {
+		cuGetErrorString(res, &(err_string));
+		rte_gpu_cuda_log(ERR,
+				"cuDeviceGetName failed with %s.\n",
+				err_string);
+
+		return -1;
+	}
+
+	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_gpu_cuda_log_debug("dev id = %u name = %s\n", 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_gpu_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_gpu_cuda_log(ERR, "Device %i failed to uninit: %i", gpu_id, ret);
+
+	rte_gpu_cuda_log_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..92b30c35b4
--- /dev/null
+++ b/drivers/gpu/cuda/meson.build
@@ -0,0 +1,13 @@
+# 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_dep = dependency('cuda', version : '>=11', modules: ['cuda'])
+ext_deps += cuda_dep
+
+deps += ['gpudev','pci','bus_pci', 'hash']
+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