Skip to content

saltzm/async_io_uring

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Overview

AsyncIOUring is an event loop that wraps the IO_Uring library with coroutines support. It supports all IO_Uring operations (with the intentional exception of poll_update*).

In addition, it allows:

  • Adding timeouts to operations
  • Manual cancellation of operations
  • Writing custom operations for advanced use cases

It is currently functionally complete, though there are a few TODOs marked in the source related to polishing the API. It's not used in production anywhere currently.

See src/async_io_uring.zig for full API documentation.

See the examples directory for an echo client and server that use the event loop.

* If you need this for some reason, please create an issue.

⚠️ The main branch of async_io_uring will follow changes to zig's master branch to stay up-to-date with changes to the IO_Uring API (among others). See the tagged releases of async_io_uring that are marked to work with specific stable versions of zig. (E.g., release v0.1.0 works with zig 0.9.1)

Table of contents


Background

As an overview for the unfamiliar, io_uring is a new-ish Linux kernel feature that allows users to enqueue requests to perform syscalls into a submission queue (e.g. a request to read from a socket) and then submit the submission queue to the kernel for processing.

When requests from the submission queue have been satisfied, the result is placed onto completion queue by the kernel. The user is able to either poll the kernel for completion queue results or block until results are available.

Zig's IO_Uring library provides a convenient interface to the kernel's io_uring functionality. The user of IO_Uring, however, still has to manually deal with submitting requests to the kernel and retrieving events from the completion queue, which can be tedious.

This library wraps the IO_Uring library by adding an event loop that handles request submission and completion, and provides an interface for each syscall that uses zig's async functionality to suspend execution of the calling code until the syscall has been completed. This lets the user write code that looks like blocking code, while still allowing for concurrency even within a single thread.

Goals

  • Minimal: Wraps the IO_Uring library in the most lightweight way possible. This means it still uses the IO_Uring data structures in many places, like for completion queue entries. There are no additional internal data structures other than the submission queue and completion queue used by the IO_Uring library. This means there's no heap allocation. It also relies entirely on kernel functionality for timeouts and cancellation.
  • Complete: You should be able to do anything with this that you could do with IO_Uring.
  • Easy to use: Because of the use of coroutines, code written with this library looks almost identical to blocking code. In addition, operation timeouts and cancellation support is integrated into the API for all operations.
  • Performant: The library does no heap allocation and there's minimal additional logic on top of suspend/resume.

Installation

This library integrates with the zigmod package manager. If you've installed zigmod, you can add a line like the following to your root_dependencies in the zig.mod file of your project and run zigmod fetch:

root_dependencies:
  - ...
  - src: git https://github.com/saltzm/async_io_uring.git

You'll then be able to include async_io_uring.zig by doing something like:

const io = @import("async_io_uring");

The examples directory is structured roughly as you might structure a project that uses async_io_uring, with a working zig.mod file and build.zig that can serve as examples.

You'll also need a Linux kernel version that supports all of the io_uring features you'd like to use. (All testing was done on version 5.13.0.)

Example usage

Echo client

Jumping right into a realistic example, the following is a snippet of code from the echo client in the examples directory:

const io = @import("async_io_uring");

pub fn run_client(ring: *AsyncIOUring) !void {
    // Make a data structure that lets us do async file I/O with the same
    // syntax as `std.debug.print`.
    var writer = try AsyncWriter.init(ring, std.io.getStdErr().handle);

    // Address of the echo server.
    const address = try net.Address.parseIp4("127.0.0.1", 3131);

    // Open a socket for connecting to the server.
    const server = try os.socket(address.any.family, os.SOCK.STREAM | os.SOCK.CLOEXEC, 0);
    defer {
        _ = ring.close(server, null, null) catch {
            std.os.exit(1);
        };
    }

    // Connect to the server.
    _ = try ring.connect(server, &address.any, address.getOsSockLen(), null, null);

    const stdin_file = std.io.getStdIn();
    const stdin_fd = stdin_file.handle;
    var input_buffer: [256]u8 = undefined;

    while (true) {
        // Prompt the user for input.
        try writer.print("Input: ", .{});

        const read_timeout = os.linux.kernel_timespec{ .tv_sec = 10, .tv_nsec = 0 };
        // Read a line from stdin with a 10 second timeout.
        // This is the more verbose API - you can also do `ring.read`.
        const read_cqe = ring.do(
            io.Read{ .fd = stdin_fd, .buffer = input_buffer[0..], .offset = input_buffer.len },
            io.Timeout{ .ts = &read_timeout, .flags = 0 },
            null,
        ) catch |err| {
            if (err == error.Cancelled) {
                try writer.print("\nTimed out waiting for input, exiting...\n", .{});
                return;
            } else return err;
        };

        const num_bytes_read = @intCast(usize, read_cqe.res);

        // Send it to the server.
        _ = try ring.send(server, input_buffer[0..num_bytes_read], 0, null, null);

        // Receive response.
        const recv_cqe = try ring.recv(server, input_buffer[0..], 0, null, null);

        const num_bytes_received = @intCast(usize, recv_cqe.res);
        try writer.print("Received: {s}\n", .{input_buffer[0..num_bytes_received]});
    }
}

Operation timeouts

AsyncIOUring supports adding timeouts to all operations. Adding a timeout to an operation causes it to be cancelled after the specified timeout, returning an error code error.Cancelled if cancellation was successful.

An example from the unit tests:

fn testReadThatTimesOut(ring: *AsyncIOUring) !void {
    var read_buffer = [_]u8{0} ** 20;

    const ts = os.linux.kernel_timespec{ .tv_sec = 0, .tv_nsec = 10000 };
    // Try to read from stdin - there won't be any input so this should
    // reliably time out.
    const read_cqe = ring.do(
        Read{ .fd = std.io.getStdIn().handle, .buffer = read_buffer[0..], .offset = 0 },
        Timeout{ .ts = &ts, .flags = 0 },
        null,
    );
    try std.testing.expectEqual(read_cqe, error.Cancelled);
}

Operation cancellation

AsyncIOUring supports cancellation for all operations. Each operation is identified by an id that is set via a maybe_id "output parameter" in all operation submission functions (e.g. read, send, etc.). This id can then be passed to AsyncIOUring.cancel to cancel that operation.

An example from the unit tests:

fn testReadThatIsCancelled(ring: *AsyncIOUring) !void {
    var read_buffer = [_]u8{0} ** 20;

    var op_id: u64 = undefined;

    // Try to read from stdin - there won't be any input so this operation should
    // reliably hang until cancellation.
    var read_frame = async ring.do(
        Read{ .fd = std.io.getStdIn().handle, .buffer = read_buffer[0..], .offset = 0 },
        null,
        &op_id,
    );

    const cancel_cqe = try ring.cancel(op_id, 0, null, null);
    // Expect that cancellation succeeded.
    try std.testing.expectEqual(cancel_cqe.res, 0);

    const read_cqe = await read_frame;
    try std.testing.expectEqual(read_cqe, error.Cancelled);
}