Async. I/O via io_uring

xNVMe supports io_uring by mapping the NVMe NVM Commands Read and Write to io_uring opcodes:

• IORING_OP_READ / IORING_OP_WRITE

  • Mapped when using xnvme_cmd_pass(..) with payload as contiguous / fixed buffers

• IORING_OP_READV / IORING_OP_WRITEV

  • Mapped when using xnvme_cmd_pass_iov(...) with payload as iovec

• IORING_OP_READ_FIXED / IORING_OP_WRITE_FIXED

  • Mapped by the io_uring_dmabuf backend when the payload is GPU memory registered as a dma-buf, see Async. I/O on GPU memory via io_uring_dmabuf

Passthru

If you are looking to do command-passthru, that is, send arbitrary user-defined NVMe commands via the Linux kernel NVMe driver, then io_uring is not the backend configuration to use. Rather, select one of:

• opts = {.be="io_uring_cmd"},NVMe passthru via io_uring big-SQE/big-CQE, see Async. I/O via io_uring_cmd

• opts = {.be="emu"},emulated async using synchronous NVMe ioctl passthru, see Async. I/O via emu

• opts = {.be="thrpool"},thread-pool based async using synchronous NVMe ioctl passthru, see Async. I/O via thrpool

Or use a user-space driver such as SPDK or libvfn.