For the time being, I decided to get better understanding of some low-level stuff related to asynchronous programming in Rust. And what can bring your more insights about how things work without inventing your own wheel. 😅

Let’s write a simple asynchronous tasks executor without any dependencies on the well-known crates like futures.

Table of content

Intro
Simple Executor
Assemble everything together
Final thoughts

Intro

ferris-scientist

Some time ago, as a maintainer of the sntpc crate which has async support, I decided to dig deeper into Rust async executors implementation. That is because the sntpc crate has 2 versions within itself:

synchronous that relies on std::net or no_std socket implementation
asynchronous one that implements exact same interface marking functions async

I found the description of several ways how we can deal with that. Usage of #[maybe_async] looks really promising and sntpc most likely would use it in upcoming releases.

I had decided to analyze a way with custom simple asynchronous executor just to provide an implementation that would use my own lightweight runtime instead of bringing dependency on tokio or async_std. Will see whether it will work or not.

Simple Executor

So, let’s dive into our journey on building our simple custom executor. Rust Async book contains an example of an executor design which contains a dependency on futures crate. I decided to not use and dig deeper by myself.

In general, my understanding for now is that the simplest executor environment can be built by following these steps:

define a container for asynchronous tasks - futures crate defines Box type for tasks. As I mentioned before, it relies on alloc crate that requires custom allocator implementation which is too much for me at this point. 😅
make a task struct that simplifies future wrapping into inner types
create an executor, that allows tasks adding and executing them, obviously

Futures container

To keep things simple, we will store all our futures on stack. In order to do that we will introduce our type:

pub struct StackBox<'a, T: ?Sized> {
    pub value: OnceCell<Pin<&'a mut T>>,
}

impl<'a, T: ?Sized> StackBox<'a, T> {
    pub fn new(value: &'a mut T) -> Self {
        let new_box = StackBox {
            value: OnceCell::new(),
        };
        new_box
            .value
            .get_or_init(|| unsafe { Pin::new_unchecked(value) });

        new_box
    }
}

Future does not implement Unpin trait, so we have to deal with some unsafe code here. The rest should be simple:

OnceCell is used, because we set up our Future only once during task creation, and it allows us to get mutable reference to an inner value without runtime borrow checking overhead with RefCell.
our future should live as long as our StackBox is alive

Task wrapper

This is a simple struct that provides a way to pass future reference and wrap it into our StackBox for further usage in our executor.

pub struct Task<'a> {
    pub name: &'a str,
    pub future: StackBoxFuture<'a>,
}

impl<'a> Task<'a> {
    pub fn new(name: &'a str, future: &'a mut impl Future<Output = ()>) -> Self {
        Self {
            name,
            future: StackBox::new(future),
        }
    }

    pub fn new_box(name: &'a str, future: StackBoxFuture<'a>) -> Self {
        Self { name, future }
    }
}

That implementation allows us to simply wrap whatever closure we want:

async fn hello() {
    println!("Hello, world!");
}

let mut fut = async { hello().await };
let task = Task::new("hello", &mut fut);

Executor

There is our simple executor:

pub struct Executor<'a> {
    tasks: [Option<Task<'a>>; 4],
    index: usize,
    pending_callback: Option<fn(&str)>,
}

Quite simple: no channels involved for receiving new tasks, just an array of tasks that should be executed and a reference to a debug callback which is called when a task in pending.

Keen eye may notice that there is a limitation:

only 4 tasks (I do not know why I decided to have only 4). We will overcome that in the future though. In a single core embedded systems you most likely know how many tasks you are going to have in your application.
you cannot add a task to such an executor “dynamically” - you have to define your tasks in the same scope you create an executor or in an outer scope. So, something like that is not possible in the current implementation

fn add_task(executor: &mut Executor) {
    let mut task = async { ... };
    // `task` has shorter lifetime than
    // an executor
    executor.spawn(&mut task);
}

// pseudocode
fn main() {
    let mut executor = Executor::new(...);
    // ...
    add_task(&mut executor);
    executor.run();
}

Technically, that should not be a problem for small embedded systems, where you usually define all tasks you want to run during initial setup:

fn main() {
    let mut executor = Executor::new(...);
    let mut logger_task = create_logger_task();
    let mut sensor_task = create_sensor_task();
    let mut cli_task = create_cli_task();

    executor.spawn(logger_task);
    executor.spawn(sensor_task);
    executor.spawn(cli_task);
    // most likely it never returns
    executor.run();
}

So, when you do not need a scheduler to conform “real-time system” requirements, that should do the trick. Also, you will not be able to add tasks from other tasks defined (sensor_task, cli_task, etc.) with that executor implementation

Anyway, that is enough for me as a starting point. 😊

Assemble everything together

There is an example in the repository that provides a showcase how it supposes to work. As in the previous section, we allocate all the task we want in executor related scope:

async fn dummy_func(data: &str) {
    let mut counter = 0usize;

    while counter != 4 {
        sleep(2);
        let now = get_timestamp_sec();
        println!("{now}: {data}");
        yield_me().await;
        counter += 1;
    }
}

let mut binding1 = async {
    dummy_func("hello").await;
};
let mut binding2 = async {
    dummy_func("world").await;
};
let mut binding3 = async {
    dummy_func("hi").await;
};
let mut binding4 = async {
    dummy_func("rust").await;
};

Since the asynchronous version of sleep is not implemented for our runtime, I just used thread sleep to avoid burning a CPU. Print some info about execution context and yield to another task. I assume that asynchronous sleep may be implemented something like that:

struct Sleep {
    start: u64
    duration: Duration,
}

impl Future for Sleep {
    type Output = ();

    fn poll(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Self::Output> {
        let now = // get current time
        if start + duration.as_sec() >= now {
            return Poll::Ready(());
        }

        cx.waker().wake_by_ref();
        Poll::Pending
    }
}

async fn sleep_me(duration: Duration) {
    let start = // get start time
    Sleep {
        start, duration
    }.await;
}

There are some points to think about though:

when we have all tasks sleeping, the executor will poll them again and again as they are yielding till sleep duration ends. So, we can think about putting them into a “sleep queue”, track the minimum time we can sleep. And if all tasks are sleeping, we can effectively release CPU and enter a low power mode, for example. Or use a Waker for that maybe?
sleep has a dependency on a somewhat that can provide us with time ticks

Let’s think about later though. 😁

That is it, after that add them all to our executor:

let _ = executor.spawn("hello", &mut binding1);
let _ = executor.spawn("world", &mut binding2);
let _ = executor.spawn("hi", &mut binding3);
let _ = executor.spawn("rust", &mut binding4);

executor.run();

And the result is what we expect:

1731508882: hello
1731508882: Task hello is pending. Waiting for the next tick...
1731508884: world
1731508884: Task world is pending. Waiting for the next tick...
1731508886: hi
1731508886: Task hi is pending. Waiting for the next tick...
1731508888: rust
# ...
1731508904: Task rust is pending. Waiting for the next tick...
1731508906: hello
1731508906: Task hello is pending. Waiting for the next tick...
1731508908: world
1731508908: Task world is pending. Waiting for the next tick...
1731508910: hi
1731508910: Task hi is pending. Waiting for the next tick...
1731508912: rust
1731508912: Task rust is pending. Waiting for the next tick...
Done!

Final thoughts

Initial implementation of our “no-dependency” executor that is suitable for no_std environment is very simple. For now, the only thing that bothers me in the above executor implementation - hardcoded number of tasks in a list.

If you like, you can experiment with the code yourself - miniloop is available.

Will see if it will be possible to use it to enhance my sntpc crate and avoid #[maybe_async] drawbacks.

What else we should analyse:

Waker operation that makes something sensible tha No-Op
Build time definition of a number of tasks in the executor

See you next time!

Intro

Simple Executor

Futures container

Task wrapper

Executor

Assemble everything together

Final thoughts

Related Posts

Rust Async Executor for Newbies - Part 2 01 Dec 2024

Multi-arch build with Podman 29 Jan 2023

Header-only library CMake install 31 Oct 2021