Driver API. Is it possible to create a memory mapping which can be accessed by the host?

Accelerated Computing CUDA CUDA Programming and Performance
#1

Hi,
I would like to use the virtual memory management API to create a contiguous virtual address range which is accessible by the host. However, I am not sure if it is possible.

I am already able to create a device memory range which can be successfully used in kernel calls. To get a host mapping, I tried the following approaches by using the device id CU_DEVICE_CPU:

Option 1: Create host memory -> CUDA_ERROR_INVALID_DEVICE

        CUmemAllocationProp prop = {};
        prop.location.type = CU_MEM_LOCATION_TYPE_DEVICE;
        prop.location.id = CU_DEVICE_CPU;
        prop.requestedHandleTypes = CU_MEM_HANDLE_TYPE_POSIX_FILE_DESCRIPTOR;
        prop.type = CU_MEM_ALLOCATION_TYPE_PINNED;
        prop.win32HandleMetaData = nullptr;

        CUDADRV_SAFE_CALL( cuMemCreate(&handle, size, &prop, 0) );

Option 2: Allow both host and device to access device memory -> CUDA_ERROR_INVALID_VALUE

        CUmemAccessDesc desc[2];
        desc[0].flags = CU_MEM_ACCESS_FLAGS_PROT_READWRITE;
        desc[0].location.id = device_;
        desc[0].location.type = CU_MEM_LOCATION_TYPE_DEVICE;
        desc[1].flags = CU_MEM_ACCESS_FLAGS_PROT_READWRITE;
        desc[1].location.id = CU_DEVICE_CPU;
        desc[1].location.type = CU_MEM_LOCATION_TYPE_DEVICE;

        CUDADRV_SAFE_CALL( cuMemSetAccess(ptr, size, &desc[0], 2) );

Are there any other options? Is it possible at all to get this working?