All Projects โ†’ japaric โ†’ no-std-async-experiments-2

japaric / no-std-async-experiments-2

Licence: Apache-2.0, MIT licenses found Licenses found Apache-2.0 LICENSE-APACHE MIT LICENSE-MIT
Cooperative multitasking (AKA async/await) on ARM Cortex-M

Programming Languages

rust
11053 projects

no-std-async-experiments-2

(For historical reasons the name says no-std but this is specifically about embedded no_std programs that target microcontrollers like ARM Cortex-M based ones)

(Part I, about (unergonomic and limited) heapless executors and waking mechanisms, can be found here; there's hardly any explanation text there though)

Status: ๐Ÿ”ฌ Proof of Concept ๐Ÿงช

The crates is this repository will not be published on crates.io; please do not depend on them as they'll not be maintained.

Goal

The goal of this experiment was to develop a cooperative scheduler that can work within a Real Time For the Masses (RTFM) application without reducing its suitability for building real time applications, that is the cooperative scheduler should not make WCET (Worst-Case Execution Time) analysis of the overall application harder to perform.

NB: Although the goal is literally "suitable for use in RTFM"; the scheduler can be used outside RTFM applications, e.g. pure cortex-m-rt applications, as shown in the examples contained in this document.

Background

Asynchronous code

As of Rust 1.39 the async fn / .await feature has been stabilized. This language feature provides the ability to write cooperative code. async fn is used to declare asynchronous functions; within an async fn one can use the .await operator to drive an asynchronous operation (e.g. another async fn) to completion in a non-blocking fashion.

// toolchain: 1.39.0

// `std::fs::File::open`
async fn open_file(path: &Path) -> File {
    // ..
}

// `std::io::Write for std::fs::File`
async fn write_to_file(file: &mut File, bytes: &[u8]) {
    // ..
}

// `std::fs::write`
async fn write_file(path: &Path, contents: &[u8]) {
    let mut f = open_file(path).await;
    write_to_file(&mut f, bytes).await;
}

Asynchronous code is meant to be executed by a (task) executor; no executor is provided in the standard library but third party crates like async-std and tokio provide multi-threaded executors. Asynchronous code to be executed by the executor is logically split into tasks; a task is basically an instance of an async fn that has been scheduled to run on the executor.

// toolchain: 1.39.0
// async-std = "1.2.0"

use async_std::task;

fn main() {
    // schedule one instance of task `foo` -- nothing is printed at this point
    task::spawn(foo());

    println!("start");
    // start task `bar` and drive it to completion
    // this puts the executor to work
    // it's implementation defined whether `foo` or `bar` runs first or
    // whether `foo` gets to run at all
    task::block_on(bar());
}

async fn foo() {
    println!("foo");
}

async fn bar() {
    println!("bar");
}
$ cargo run
start
foo
bar

The executor executes tasks cooperatively; in its simplest form the executor will run a task until it reaches an .await operation; if that operation would need to block (e.g. because it's waiting on a socket, etc.) then executor suspends that task and moves to, resumes, another one. Thus, .await operators are potential suspension points within asynchronous code. An efficient executor will only resume tasks that it knows will be able to make progress (that is, they will not immediately yield again); this minimizes the amount of context switching between tasks.

Some implementation details

In Rust, the core building block for cooperative multitasking (AKA asynchronous code) are generators. Syntactically, a generator looks like a closure (|| { .. } ) with suspension points (yield) in it. Semantically, a generator is a state machine where each state consists of the execution of arbitrary code (between yield points) and transitions are controlled externally (using the resume method).

// toolchain: nightly-2019-12-02

use core::{pin::Pin, ops::Generator};

fn main() {
    let mut g = || {
        println!("A");
        yield;
        println!("B");
        yield;
        println!("C");
    };

    let mut state = Pin::new(&mut g).resume();
    println!("{:?}", state);
    state = Pin::new(&mut g).resume();
    println!("{:?}", state);
    state = Pin::new(&mut g).resume();
    println!("{:?}", state);
}

In its simplest form (ignoring fairness and efficiency), an executor needs to keep a list of these state machines and continuously resume them until they complete. As each of these state machines may have a different size and runs different code when resume-d some form of indirection is required to store them in a list. So each element in the list will be a trait object, e.g. Box<dyn Generator>, instead of a concrete generator, i.e. impl Generator.

fn executor(mut tasks: Vec<Pin<Box<dyn Generator<Yield = (), Return = ()>>>>) {
    let mut n = tasks.len();
    while n != 0 {
        for i in (0..n).rev() {
            let state = tasks[i].as_mut().resume();
            if let GeneratorState::Complete(()) = state {
                tasks.swap_remove(i); // done; remove
            }
        }

        n = tasks.len();
    }
}

Idea

The approach that will be explored here will consist of isolating all cooperative multitasking ("asynchronous code") to the #[idle] context (or fn main in non-RTFM applications).

There are few reasons for this:

  • I expect that some of cooperative tasks that users will write will be never-ending and some others will be short-lived. Thus it's sensible to run the executor in #[idle], which is the never-ending background context.

  • All the dynamic memory allocations required by the executor can be constrained to #[idle]. Meaning that we can use a non-real-time allocator in #[idle] and leave regular tasks completely free of dynamic memory allocation (*). As the allocator will be exclusively used in #[idle] we don't need any form of mutex to protect it -- #[idle] effectively owns the allocator.

(There's an hypothetical third reason: the possibility of hyper-tuning the allocator; this will be explored later on)

(*) Or least make dynamic allocation in tasks opt-in. We can give tasks a resource-locked TLSF allocator; this gives them the ability to alloc and dealloc (but not realloc) in bounded constant time regardless of the size of the allocation.

Implementation

The implementation has two main components: a "thread-mode" allocator and a "thread-mode" executor.

The "thread-mode" moniker is a bit unfortunate but it refers to the fact that the allocator / executor will be constrained to what ARM calls "Thread mode". That is the allocator / executor can not be accessed / used from "Handler mode" (another ARM term). "Handler mode" is basically interrupt / exception context, whereas "Thread mode" is non-interrupt context. All code executed by the Reset handler (e.g. after booting) runs in "Thread mode"; in RTFM apps #[init] and #[idle] run in "Thread mode"; in cortex-m-rt apps #[entry] runs in "Thread mode".

TM (Thread-Mode) allocator

(NB: a complete version of the snippets presented here can be found in the examples directory. You can run the examples (cargo run) if you have QEMU and the thumb7m-none-eabi installed; you can find installation instructions in [the embedded Rust book])

The TM allocator is a separate allocator, independent of the global allocator one can be define with #[global_allocator]. Ideally, we would use the Alloc trait and allocator-generic collections specified (the later loosely specified) in RFC #1398 to implement this allocator but the former, the Alloc trait, is unstable and the later, the collections, don't exist -- all alloc collections are hard-coded to use the #[global_allocator].

On the bright side, we can implement the TM allocator on stable but we can't implement all the collections because some of them depend on unstable features (e.g. core::intrinsics::abort in Rc). Also, on stable we can't implement coercion for these collections so you can't coerce Box<impl Generator> (concrete type) into Box<dyn Generator> (trait object) or Box<[u8; 64]> (array) into Box<[u8]> (slice).

The API devised for the TM allocator looks like this:

// toolchain: 1.39.0

use cortex_m_tm_alloc::allocator;
use tlsf::Tlsf;

#[allocator(lazy)]
static mut A: Tlsf = {
    // `MEMORY` is transformed into `&'static mut [u8; 64]`
    static mut MEMORY: [u8; 64] = [0; 64];

    let mut tlsf = Tlsf::new();
    tlsf.extend(MEMORY);
    tlsf
};

This defines a TM allocator named A. A is a lazy-ly (runtime) initialized TLSF allocator. The block expression on the right hand side of the static item is executed at runtime exactly once, so the usual static mut X: T -> X: &'static mut T transformation applies. Also, since this code is executed at runtime rather than at compile time the allocator constructor doesn't need to be a const fn.

You can get a handle to the A allocator using the get constructor. The constructor returns Option<A>; when called in "thread-mode" it always returns the Some variant; on the other hand it always returns None when called in "handler-mode". A is a zero sized type (ZST) that implements the Copy and Alloc traits but doesn't implement the Send and Sync traits. Not implementing Send ensures an instance is never sent into an interrupt / exception handler.

#[entry]
fn main() -> ! {
    hprintln!("before A::get()").ok();
    SCB::set_pendsv();

    if let Some(a) = A::get() {
        hprintln!("after A::get()").ok();
        SCB::set_pendsv();

        // ..
    } else {
        // UNREACHABLE
    }

    // ..
}

#[exception]
fn PendSV() {
    hprintln!("PendSV({:?})", A::get()).ok();
}
$ cargo run
before A::get()
PendSV(None)
after A::get()
PendSV(None)

Once you get an instance of A you can use it initialize to a collection by passing a copy of it to its constructor. (The collection stores a copy of the allocator handle; thanks to A being a ZST this doesn't increase the stack-size of the collection)

if let Some(a) = A::get() {
    // ..

    let mut xs: Vec<i32, A> = Vec::new(a);

    for i in 0.. {
        xs.push(i);
        hprintln!("{:?}", xs).ok();
    }
}

As with the global allocator, a TM allocator may run out of memory. In that case, the Out-Of-Memory handler defined using the #[oom] attribute will get called.

#[alloc_oom::oom]
fn oom(layout: Layout) -> ! {
    hprintln!("oom({:?})", layout).ok();
    debug::exit(debug::EXIT_FAILURE);
    loop {}
}
$ cargo run
[0]
[0, 1]
[0, 1, 2]
[0, 1, 2, 3]
oom(Layout { size_: 32, align_: 4 })

$ echo $?
1

It should be noted that TM allocators never mask interrupts; they don't internally use RefCell either so no runtime checks or panicking branches there; the fast path of their get constructor compiles down to 3 instructions (load, shift left, conditional branch); and their lazy initialization uses a single extra byte of static memory and doesn't require atomics.

TM (Thread-Mode) executor

The other component is the TM (task) executor. This executor depends on a TM allocator and it's declared like this:

// toolchain: nightly-2019-12-02

use cortex_m_tm_alloc::allocator;
use cortex_m_tm_executor::executor;

#[allocator(lazy)]
static mut A: Tlsf = { /* .. */ };

executor!(name = X, allocator = A);

Like the TM allocator you can only get a handle to this TM executor in thread-mode context using the get constructor. This functions returns a handle to the TM executor and a handle to its TM allocator. The handle to the TM executor is Copy but not Send or Sync.

The executor handle can be used to spawn tasks. spawn takes a concrete generator, boxes it and stores it in an internal queue; spawn doesn't execute any of the generator / task code! To start executing tasks you use the block_on API. This function takes a generator that will driven to completion (without boxing it) while making progress on other previously spawned tasks.

Here's an example:

#[entry]
fn main() -> ! {
    if let Some((x, _a)) = X::get() {
        x.spawn(move || {
            hprintln!(" A0").ok();
            yield;

            hprintln!(" A1").ok();
            // but of course you can `spawn` a task from a spawned task
            x.spawn(|| {
                hprintln!("  C0").ok();
                yield;

                hprintln!("  C1").ok();
            });
            yield;

            hprintln!(" A2").ok();
            // NOTE return value will be discarded
            42
        });

        let ans = x.block_on(|| {
            hprintln!("B0").ok();
            yield;

            hprintln!("B1").ok();
            yield;

            hprintln!("B2").ok();
            yield;

            hprintln!("B3").ok();

            42
        });

        hprintln!("the answer is {}", ans).ok();
    }

    debug::exit(debug::EXIT_SUCCESS);

    loop {}
}

Given that the executor handle is Copy one can move the handle into a spawned task and spawn another task from it. This can be seen in the example: main spawns task A and task A spawns task C.

block_on does not guarantee that all previously spawned tasks will be driven to completion; it only drives its argument generator to completion. This can be seen in the example: task C is not completed by the time block_on returns.

You shall (not?) deadlock

block_on seems better than spawn because it doesn't box its generator and it's able to preserve the return value of the generator. However, nesting block_on calls can lead to deadlocks. Here's an example using async-std.

// toolchain: 1.39.0
// async-std = "1.2.0"

use async_std::{sync, task};

fn main() {
    task::block_on(async {
        let (s, r) = sync::channel::<i32>(1);

        task::block_on(async move {
            let x = r.recv().await;
            println!("got {:?}", x);
        });

        s.send(0).await;
        println!("send");
    });
}

This program hangs and nothing is printed to the console. If we replace the inner block_on with spawn then we get the intended output of:

$ cargo run
send
got Some(0)

(I know, I know. "Nobody writes code like this". I agree this is unlikely to occur if the program is short enough to fit in a single file but I think the chances of running into are non-negligible once your program spans several files, or worst crates)

For this reason and to simplify the implementation nesting block_on calls will panic the TM executor. (I think it's not possible to deadlock the executor with this restriction but have no way to prove it)

#[r#async] / r#await!

If you can't nest block_on calls then how do we drive generators to completion (in a non-blocking fashion)? We use the r#await! macro. There's also an #[r#async] attribute that save typing time when writing functions that return generators.

As an example, let's say you want to asynchronous receive items send from an interrupt handler. You would write your asynchronous function using #[r#async] like this:

use core::ops::Generator;

use heapless::{
    spsc::Consumer, // consumer endpoint of a single-producer single-consumer queue
    ArrayLength,
};
use gen_async_await::r#async;

#[r#async]
fn dequeue<T, N>(mut c: Consumer<'static, T, N>) -> (T, Consumer<'static, T, N>)
where
    N: ArrayLength<T>,
{
    loop {
        if let Some(x) = c.dequeue() {
            break (x, c);
        }
        yield
    }
}

// OR you could have written this; both are equivalent
fn dequeue2<T, N>(
    mut c: Consumer<'static, T, N>,
) -> impl Generator<Yield = (), Return = (T, Consumer<'static, T, N>)>
where
    N: ArrayLength<T>,
{
    || loop {
        if let Some(x) = c.dequeue() {
            break (x, c);
        }
        yield
    }
}

(If you are wondering why I'm passing the Consumer by value rather than by reference: it's to work around the lack of support for self-referential generators; I'll get back to this later on)

Then in the application you would write something like this:

#[entry]
fn main() -> ! {
    static mut Q: Queue<i32, consts::U4> = Queue(i::Queue::new());

    let (p, mut c) = Q.split();

    // send the producer to an interrupt handler
    send(p);

    if let Some((x, _a)) = X::get() {
        // task that asynchronously processes items produced by
        // the interrupt handler
        x.spawn(move || loop {
            let ret = r#await!(dequeue(c)); // <-
            let item = ret.0;
            c = ret.1;
            // do stuff with `item`
        });

        x.block_on(|| {
            // .. do something else ..
        });
    }

    debug::exit(debug::EXIT_SUCCESS);

    loop {}
}

The infinite task will r#await! the #[r#async] function we wrote before.

Fixed capacity queue

The TM executor uses a variable capacity queue by default (e.g. alloc::Vec) but it's possible to switch to a fixed capacity queue (e.g. heapless::Vec). If you can upper bound the number of spawned tasks in your program it may be advantageous to use a fixed capacity queue. With a fixed capacity queue, the queue is allocated once, and could even be allocated on the stack. Plus, if you are using the allocator only for the task executor then the compiler can optimize away the realloc routine (and the grow_in_place and shrink_in_place routines called by it) as only alloc and dealloc are required to box generators and destroy them.

The syntax to switch to the fixed capacity queue is shown below:

use executor::executor;

// fixed-capacity = 4 tasks
executor!(name = X, allocator = A, max_spawned_tasks = U4);

std_async::sync::Mutex?

With async-std if you want to share memory between two tasks you need to reach out for its Mutex or its RwLock abstraction because the task::spawn API requires that its argument generator (and all its captures) implement the Send trait. This is required because the executor is multi-threaded and tasks could run in parallel.

use async_std::{sync::Mutex, task};

fn main() {
    let shared: &'static Mutex<u128> = Box::leak(Box::new(Mutex::new(0u128)));

    task::spawn(async move {
        let x = shared.lock().await;
        println!("{}", x);
    });

    task::block_on(async move {
        *shared.lock().await += 1;
    });
}

In our case, the TM executor runs everything on the same context so tasks will always be resumed serially (one after the one) and without overlap. Thus no Send bound is required on the generator passed to spawn; therefore instead of a Mutex you use a plain RefCell (or a Cell) to share data between tasks.

use core::cell::RefCell;

#[entry]
fn main() -> ! {
    static mut SHARED: RefCell<u64> = RefCell::new(0);

    if let Some((x, _a)) = X::get() {
        let shared: &'static _ = SHARED;

        x.spawn(move || loop {
            hprintln!("{}", shared.borrow()).ok();
            yield;
        });

        x.block_on(move || {
            *shared.borrow_mut() += 1;
            yield;
            *shared.borrow_mut() += 1;
            yield;
        });
    }

    debug::exit(debug::EXIT_SUCCESS);

    loop {}
}

Note that it's not necessary to r#await! to access the shared data.

std::sync::Arc?

Sometimes you want the shared data to eventually be freed up. In async-std you would reach out for a Send-able Arc to delegate the destruction of the data to the last user.

use async_std::{sync::Arc, task};

fn main() {
    let a = Arc::new(0);
    let b = a.clone();
    task::spawn(async move {
        if Arc::strong_count(&a) == 1 {
            println!("A: I'll destroy the Arc!")
        }
        drop(a);
    });

    task::spawn(async move {
        if Arc::strong_count(&b) == 1 {
            println!("B: I'll destroy the Arc!")
        }
        drop(b);
    });

    task::block_on(async move {
        // do something else
    });
}

With the TM executor you can make do with a non-atomic Rc.

#[entry]
fn main() -> ! {
    if let Some((x, a)) = X::get() {
        let a = Rc::new(0, a);
        let b = a.clone();

        x.spawn(move || {
            if Rc::strong_count(&a) == 1 {
                hprintln!("A: I'll destroy the Rc!").ok();
            }
            drop(a);
            yield;
        });

        x.spawn(move || {
            if Rc::strong_count(&b) == 1 {
                hprintln!("B: I'll destroy the Rc!").ok();
            }
            drop(b);
            yield;
        });

        x.block_on(move || {
            yield;
            yield;
        });
    }

    debug::exit(debug::EXIT_SUCCESS);

    loop {}
}

Limitations

Self-referential generators

AFAICT, as of nightly-2019-12-02, all generators created using the || { yield } syntax are marked Unpin (they are movable) and self-referential borrows are forbidden inside them. See below:

fn main() {
    let g = || {
        let x = 0;
        let y = &x; //~ ERROR: borrow may still be in use when generator yields
        yield; //~ INFO: possible yield occurs here
        drop(y);
    };
}

This limits what one can do with #[r#async] / r#await! implemented directly on top of generators. In contrast, in the built-in async fn / .await language feature the future returned by an async fn function is always marked !Unpin (the future is immovable) and self-referential borrows are allowed. See below:

use core::{future::Future, ops::Generator};

fn main() {
    let g = || yield;
    is_future(&g); // always false
    is_generator(&g); // always true
    is_unpin(&g); // (currently?) always true

    let f = foo();
    is_future(&f); // always true
    is_generator(&f); // always false
    is_unpin(&f); // (currently?) always false
}

async fn foo() {}

fn is_future(_: &impl Future) {}
fn is_generator(_: &impl Generator) {}
fn is_unpin(_: &impl Unpin) {}

We saw an example of what can't be written due to the lack self-referential generators in the dequeue function (section #[r#async] and r#await!). Here I show that the function can be written in the intended way using futures and async fn / .await.

#[r#async] / r#await! version

// so, actually you *can* write this
#[r#async]
fn dequeue<'a, T, N>(c: &'a mut Consumer<'static, T, N>) -> T
where
    N: ArrayLength<T>,
{
    loop {
        if let Some(x) = c.dequeue() {
            break x;
        }
        yield
    }
}

// but then you cannot use it :-(
#[r#async]
fn task(mut c: Consumer<'static, i32, U4>) {
    loop {
        //~ ERROR: borrow may still be in use when generator yields
        let item = r#await!(dequeue(&mut c));
        //~ INFO:                   ^^^^^^ possible yield occurs here
        // do stuff
    }
}

async fn / .await version

use core::{future::Future, ops::Generator, task::Poll};

use futures::future;
use heapless::{spsc::Consumer, ArrayLength, consts::U4};

fn dequeue<'a, T, N>(c: &'a mut Consumer<'static, T, N>) -> impl Future<Output = T> + 'a
where
    N: ArrayLength<T>,
{
    future::poll_fn(move |cx| {
        if let Some(item) = c.dequeue() {
            return Poll::Ready(item);
        }
        // NOTE: dumb but, without this, awaiting this future may hang the thread
        cx.waker().wake_by_ref();
        Poll::Pending
    })
}

async fn task(mut c: Consumer<'static, i32, U4>) {
    loop {
        let item = dequeue(&mut c).await;
    }
}

What's even this Unpin stuff?

The Pin abstraction and the Unpin marker trait enable the memory-safe creation and use of structs with self-referential fields. Self-referential fields mean that one of the field of the struct can be a reference to a another field of the same struct.

"Wait, what do structs have to do with generators?" The generator syntax is sugar for creating a struct that implements the Generator trait; the state of the generator is stored in the fields of this struct.

For example the following generator:

(which as I mentioned is not accepted today but I hope it'll be accepted in the future and marked !Unpin)

let g = || {
    let x = 0;
    let y = &x;
    yield;
    use(y);
};

could be represented by the following struct:

struct G {
    x: i32,
    // (`'self` is made up syntax)
    y: &'self i32, // points to `x`

    state: i32, // keeps track of which `yield` where are currently at
}

x is part of the state of the generator so it's a field of the struct. y is a reference to this variable but also part of the state so it needs to be a field too. Thus the struct contains a self-referential reference: y points to the field x.

It would be problematic if we could resume this generator and then move it. Here's an example:

fn foo() {
   let mut g = bar();
   Pin::new(&mut g).resume();
}

fn bar() -> impl Generator {
    let mut g = || {
        let x = 0;
        let y = &x;
        yield;
        use(y);
    };

    Pin::new(&mut g).resume();
    g
}

In bar we create the generator and resume it once. At that point, the state of the generator contains initialized x and y fields. However, both x and y live on the stack frame of function bar and y points into the stack frame of bar. When the generator is returned from bar to foo the stack frame where the generator was created is destroyed. By the time we call the second resume in foo, x is valid but y is not; it's still pointing into the destroyed stack frame. This second resume call is UB.

The way Rust prevents this misuse of generators is the Unpin marker trait. One can only call Pin::new(&mut g) if g implements the Unpin trait. However, in this case the generator is !Unpin because of the self-referential borrow. Thus the compiler rejects this code at compile time.

Does this means we can't never resume generators that contain self-referential borrows? It's possible to resume them but they need to be properly pinned. This updated example compiles and its sound:

fn foo() {
   let mut bg = bar();
   bg.resume();
}

fn bar() -> Box<Pin<impl Generator>> {
    let g = || {
        let x = 0;
        let y = &x;
        yield;
        use(y);
    };
    let mut bg = Box::pin(g);

    bg.resume();
    bg
}

In the updated example we first box-pin the generator. The boxed generator can be safely resumed in bar and in foo. The reason this operation is now safe is that bg.x is stored in the heap and has a stable address. Moving the boxed generator from bar to foo doesn't change the address of bg.x so the self-reference bg.y is not invalidated by the move.

Potential improvements

Hyper-tuning the allocator

As I mentioned before if one goes down the route of using the TM allocator only in the TM executor and bounds the maximum number of spawned tasks then the TM allocator only needs to alloc and dealloc generators (tasks). As all these generators are : Sized we can extract their size information from the output of the compiler (e.g. LLVM IR or machine code). With this information, we could optimize the allocator for this particular payload; this could either mean using just a few single-linked lists of free blocks as the allocator (fastest option, constant time even, but not memory efficient) or configuring a more general allocator to have size classes that closely match the expected allocation requests.

(More information about allocator strategies and trade-offs can be found in the paper "Dynamic Storage Allocation: A Survey and Critical Review")

Less dumb executor

The executor in this proof of concept is extremely dumb: it resumes all tasks non-stop (without ever sleeping) in a round-robin-ish fashion until they complete.

In contrast, async fn / .await executors only resume tasks that can make progress (that have been "woken"). This maximizes throughput in multi-threaded systems but I'm not sure the complex book-keeping required to achieve this "perfect resumption" is worth the effort in constrained systems that will only handle a few concurrent tasks. The CPU time spend in book-keeping could be comparable to the amount of work spent in application logic (some embedded systems do little work and sleep most of the time). Plus, the memory used for book-keeping could be a sizable part of the limited amount found in these systems. More worrying though is that HAL authors will likely be writing this complex book-keeping code to provide async APIs in their libraries; the more complex the code to integrate the higher chance to also introduce bugs.

Any improvement on this front needs to take all these factors into account: being able to get some sleep is better than perfectly avoiding any unnecessary resumptions; book-keeping needs to be lightweight, both in terms of memory usage and CPU time; the solution needs to be easy to integrate by HAL authors.

Other observations

Panicking branches

Resuming even the simplest generator || { yield } produces machine code that contains panicking branches that will never be reached at runtime. This may be solved by improving the MIR / LLVM-IR codegen pass in rustc. More details can be found in rust-lang/rust#66100.

Benchmarks

The context switching performance of the proof of concept was measured using the following snippet:

x.spawn(move || {
    asm::bkpt();
    yield;
});

x.spawn(|| {
    asm::bkpt();
    yield;
});

x.block_on(|| {
    asm::bkpt();
    yield;

    asm::bkpt();
    yield;
});

The CYCcle CouNTer (CYCCNT) was read at each breakpoint and the difference was found to be 19-36 cycles.

As a reference the context switching in an RTFM application was also measured using the following snippets.

  • Software tasks
#[task(priority = 2)]
fn a(cx: a::Context) {
    asm::bkpt();
}

#[task(priority = 2)]
fn b(cx: b::Context) {
    asm::bkpt();
}

#[task(priority = 1)]
fn c(cx: c::Context) {
    asm::bkpt();
}

a -> b (same priority) took 24 cycles; b -> c (to lower priority) took 35 cycles.

  • Hardware tasks
#[task(binds = EXTI0, priority = 2)]
fn a(cx: a::Context) {
    asm::bkpt();
}

#[task(binds = EXTI1, priority = 2)]
fn b(cx: b::Context) {
    asm::bkpt();
}

#[task(binds = EXTI2, priority = 1)]
fn c(cx: c::Context) {
    asm::bkpt();
}

Both a -> b and b -> c took 8 cycles.

Conclusions

We have presented a proof of concept implementation of async-await that can be used today in embedded no-std code (though it requires nightly).

The cooperative scheduler (AKA executor) is completely isolated from interrupt handlers. Thus the cooperative code can run without compromising the predictability of real-time / time-sensitive code running in interrupt handlers. This has been achieved by having the executor use a dedicated memory allocator, separate from the #[global_allocator]. Furthermore, in some configurations one can extract information from the compiler to optimize this allocator for the target application.

Lastly, the memory allocator API developed during this experiment can be used in no-std binary crates on stable, unlike the #[global_allocator] API which depends on the unstable #[alloc_error_handler] feature. Unfortunately, this allocator API does not inter-operate with the collections in alloc so these collections need to be re-implemented to use the alternative allocator API.

License

Licensed under either of

at your option.

Contribution

Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in the work by you, as defined in the Apache-2.0 license, shall be dual licensed as above, without any additional terms or conditions.

Note that the project description data, including the texts, logos, images, and/or trademarks, for each open source project belongs to its rightful owner. If you wish to add or remove any projects, please contact us at [email protected].