Is it legal to pass the dma address of a GPU buffer to the Ib verbs for Memory Registration for initiating RDMA. Does MR differentiate between device memory/cpu memory.
In my NVME-OF test setup, I do a 128K read over fabric from a file to a GPU device memory directly. The nvme rdma driver at host gets a completion status OK. But on checking the actual GPU memory content, it seems the read did not go through. Is there a way to debug the RDMA issue.