I am currently working on GPUDirect RDMA implementation on a proprietary interconnect, which accepts virtual addresses only. I have some questions on GPUDirect RDMA functionalities (https://docs.nvidia.com/cuda/gpudirect-rdma/index.html):
- Is it possible to access directly the GPU memory without involving get_user_pages() ?
- What exactly get_user_pages() call does ? It only maps the GPU buffer on the PCI BAR1 ? Will this call add an entry to the page table? Is it possible to access the GPU page table (virtual to physical address translation)?
My Objective: Be able to pass a virtual address to the NIC which corresponds to a GPU physical memory address.