Skip to content

Zero-Syscall Kernel I/O via io_uring (SQPOLL)

Ship provides raw, bare-metal built-in function wrappers for the Linux io_uring subsystems and memory mapping system calls. This enables high-performance, asynchronous, zero-syscall user-space I/O.

1. Exposed Compiler Built-ins

The following low-level built-in primitives are now wired directly into the compiler, lowering to their respective Linux x86_64 system calls:

sys_io_uring_setup

Initializes a new io_uring instance. 

fn sys_io_uring_setup(entries int, params int) -> int
* Syscall ID: 425 * Parameters: * entries: Size of the submission and completion queues. * params: Memory pointer to the io_uring_params structural setup. * Returns: File descriptor of the io_uring instance, or -1 on error.

sys_io_uring_enter

Initiates and completes I/O operations. 

fn sys_io_uring_enter(fd int, to_submit int, min_complete int, flags int) -> int
* Syscall ID: 426 * Parameters: * fd: io_uring ring file descriptor. * to_submit: Number of entries to submit. * min_complete: Minimum completed operations to wait for. * flags: Completion modifiers (e.g. IORING_ENTER_GETEVENTS).

sys_io_uring_register

Registers user-space buffers or files for kernel polling optimization. 

fn sys_io_uring_register(fd int, opcode int, arg int, nr_args int) -> int
* Syscall ID: 427

sys_mmap

Maps memory pages shared between user space and kernel space. 

fn sys_mmap(addr int, len int, prot int, flags int, fd int, offset int) -> int
* Syscall ID: 9

2. Implementing Zero-Syscall I/O (SQPOLL)

To bypass the context-switching penalty of system calls entirely:

  1. Enable SQPOLL Mode: Set the IORING_SETUP_SQPOLL flag inside your io_uring_params struct before calling sys_io_uring_setup.
  2. Kernel Thread Loop: The kernel spawns a dedicated background thread that continuously polls the shared Submission Queue (SQ).
  3. Queue Updates: In user-space, write requests (e.g., file reads, packet sends) straight to the submission ring memory mapped via sys_mmap, and increment the queue's tail pointer.
  4. Result Collection: Read results from the Completion Queue (CQ) buffers.

This lock-less ring buffer flow entirely avoids syscall transitions, context switches, and TLB flushes, achieving maximum hardware throughput.