rio 0.9.4

GPL-3.0 nice bindings for io_uring. MIT/Apache-2.0 license is available for spacejam's github sponsors.
Documentation

rio

bindings for io_uring, the hottest thing to happen to linux IO in a long time.

Soundness status

rio aims to leverage Rust's compile-time checks to be misuse-resistant compared to io_uring interfaces in other languages, but users should beware that use-after-free bugs are still possible without unsafe when using rio. Completion borrows the buffers involved in a request, and its destructor blocks in order to delay the freeing of those buffers until the corresponding request has completed; but it is considered safe in Rust for an object's lifetime and borrows to end without its destructor running, and this can happen in various ways, including through std::mem::forget. Be careful not to let completions leak in this way, and if Rust's soundness guarantees are important to you, you may want to avoid this crate.

Innovations

  • only relies on libc, no need for c/bindgen to complicate things, nobody wants that
  • the completions work great with threads or an async runtime (Completion implements Future)
  • uses Rust marker traits to guarantee that a buffer will never be written into unless it is writable memory. (prevents you from trying to write data into static read-only memory)
  • no need to mess with IoSlice / libc::iovec directly. rio maintains these in the background for you.
  • If left to its own devices, io_uring will allow you to submit more IO operations than would actually fit in the completion queue, allowing completions to be dropped and causing leaks of any userspace thing waiting for the completion. rio exerts backpressure on submitters when the number of in-flight requests reaches this threshold, to guarantee that no completions will be dropped due to completion queue overflow.
  • rio will handle submission queue submissions automatically. If you start waiting for a Completion, rio will make sure that we have already submitted at least this request to the kernel. Other io_uring libraries force you to handle this manually, which is another possible source of misuse.

This is intended to be the core of sled's writepath. It is built with a specific high-level application in mind: a high performance storage engine and replication system.

What's io_uring?

io_uring is the biggest thing to happen to the linux kernel in a very long time. It will change everything. Anything that uses epoll right now will be rewritten to use io_uring if it wants to stay relevant. It started as a way to do real async disk IO without needing to use O_DIRECT, but its scope has expanded and it will continue to support more and more kernel functionality over time due to its ability to batch large numbers different syscalls. In kernel 5.5 support is added for more networking operations like accept(2), sendmsg(2), and recvmsg(2). In 5.6 support is being added for recv(2) and send(2). io_uring has been measured to dramatically outperform epoll-based networking, with io_uring outperforming epoll-based setups more and more under heavier load. I started rio to gain an early deep understanding of this amazing new interface, so that I could use it ASAP and responsibly with sled.

io_uring unlocks the following kernel features:

  • fully-async interface for a growing number of syscalls
  • async disk IO without using O_DIRECT as you have to do with AIO
  • batching hundreds of disk and network IO operations into a single syscall, which is especially wonderful in a post-meltdown/spectre world where our syscalls have dramatically slowed down
  • 0-syscall IO operation submission, if configured in SQPOLL mode
  • configurable completion polling for trading CPU for low latency
  • Allows expression of sophisticated 0-copy broadcast semantics, similar to splice(2) or sendfile(2) but working with many file-like objects without ever needing to bounce memory and mappings into userspace en-route.
  • Allows IO buffers and file descriptors to be registered for cheap reuse (remapping buffers and file descriptors for use in the kernel has a significant cost).

To read more about io_uring, check out:

For some slides with interesting io_uring performance results, check out slides 43-53 of this presentation deck by Jens.

why not use those other Rust io_uring libraries?

  • they haven't copied rio's features yet, which you pretty much have to use anyway to responsibly use io_uring due to the sharp edges of the API. All of the libraries I've seen as of January 13 2020 are totally easy to overflow the completion queue with, as well as easy to express use-after-frees with, don't seem to be async-friendly, etc...

examples that will be broken in the next day or two

async tcp echo server:

use std::{
    io::self,
    net::{TcpListener, TcpStream},
};

async fn proxy(ring: &rio::Rio, a: &TcpStream, b: &TcpStream) -> io::Result<()> {
    let buf = vec![0_u8; 512];
    loop {
        let read_bytes = ring.read_at(a, &buf, 0).await?;
        let buf = &buf[..read_bytes];
        ring.write_at(b, &buf, 0).await?;
    }
}

fn main() -> io::Result<()> {
    let ring = rio::new()?;
    let acceptor = TcpListener::bind("127.0.0.1:6666")?;

    extreme::run(async {
        // kernel 5.5 and later support TCP accept
        loop {
            let stream = ring.accept(&acceptor).await?;
            dbg!(proxy(&ring, &stream, &stream).await);
        }
    })
}

file reading:

let ring = rio::new().expect("create uring");
let file = std::fs::open("file").expect("openat");
let data: &mut [u8] = &mut [0; 66];
let completion = ring.read_at(&file, &mut data, at);

// if using threads
completion.wait()?;

// if using async
completion.await?

file writing:

let ring = rio::new().expect("create uring");
let file = std::fs::create("file").expect("openat");
let to_write: &[u8] = &[6; 66];
let completion = ring.write_at(&file, &to_write, at);

// if using threads
completion.wait()?;

// if using async
completion.await?

speedy O_DIRECT shi0t (try this at home / run the o_direct example)

use std::{
    fs::OpenOptions, io::Result,
    os::unix::fs::OpenOptionsExt,
};

const CHUNK_SIZE: u64 = 4096 * 256;

// `O_DIRECT` requires all reads and writes
// to be aligned to the block device's block
// size. 4096 might not be the best, or even
// a valid one, for yours!
#[repr(align(4096))]
struct Aligned([u8; CHUNK_SIZE as usize]);

fn main() -> Result<()> {
    // start the ring
    let ring = rio::new()?;

    // open output file, with `O_DIRECT` set
    let file = OpenOptions::new()
        .read(true)
        .write(true)
        .create(true)
        .truncate(true)
        .custom_flags(libc::O_DIRECT)
        .open("file")?;

    let out_buf = Aligned([42; CHUNK_SIZE as usize]);
    let out_slice: &[u8] = &out_buf.0;

    let in_buf = Aligned([42; CHUNK_SIZE as usize]);
    let in_slice: &[u8] = &in_buf.0;

    let mut completions = vec![];

    for i in 0..(10 * 1024) {
        let at = i * CHUNK_SIZE;

        // By setting the `Link` order,
        // we specify that the following
        // read should happen after this
        // write.
        let write = ring.write_at_ordered(
            &file,
            &out_slice,
            at,
            rio::Ordering::Link,
        );
        completions.push(write);

        let read = ring.read_at(&file, &in_slice, at);
        completions.push(read);
    }

    for completion in completions.into_iter() {
        completion.wait()?;
    }

    Ok(())
}