Trying to bring reusable and efficient components to embedded Rust has been a challenge for our team. We think we've started to make headway, and want to introduce the Drogue Device project.

Background

A lot of embedded frameworks and RTOSs rely on procedural coding, "copy these source files into your tree" or a lot of manual connection of APIs to the runtime. Since We Are Red Hat (or IBM, or more specifically/historically "JBoss"), we like reusable components. And frameworks. They represent guardrails and provide guidance on how to build larger systems, we feel.

Component Systems

In enterprise software, component systems abound. There's Enterprise Java Beans, Actix, CDI, Vert.x, Node.js, Akka framework, Quarkus, Spring, etc. Many of these component systems tend to rely on multi-threading capability of the underlying system, and bring about concurrency issues, such as shared-state and locking.

Actor Systems

Of these systems, Akka is an actor system, which attempts to draw clear lines about ownership and control of state and data, helping alleviate concurrency concerns. An actor within an actor system believes that it is single-threaded, serially processing requests that arrive which desire to operate upon the data it controls. This seemingly single-threaded nature allows an actor to avoid locking of its data. All mutations occur from the POV of the actor himself, triggered by message passing from external sources which make requests.

Messaging

Common messaging patterns include fire-and-forget notifications from one component to another, and request/response between two components. Sometimes components want to ambiguously broadcast messages to whichever unknown components might be interested. This third type is basically an event bus.

Task Scheduling

An actor that thinks it is single-threaded is great, but our MCUs usually also only have one core, and thus one thread. A slow actor, or one that blocks the thread of execution can bring the entire system to a halt. Just because one actor hits a wall and can make no further progress, another actor might be able to do some meaningful work, unblocking the first.

Preemptive Task Scheduling

If you have a preemptive thread or task scheduler, then a kernel is slicing up processor time. An task is given a certain number of cycles (or milliseconds, or reductions or whatever unit measurement used to slice the time pie), and when the timer expires, the kernel forcibly pause the task and select another one to run for a slice of time. Naive implementations will allow a blocked task to continue to block until his time-slice expires. Other implementations allow a task to indicate that it is blocked, and relinquish any remaining time, cooperatively.

Run-to-Completion Task Scheduling

Some schedulers select a task, and allow it to run to completion, even if it's slow, or blocks briefly. The onus is on the developer to ensure his tasks will not completely halt or take too much time. Run-to-completion semantics can be great when you have a task you absolutely need to have, well, run to completion before being put into the queue to run again later.

Cooperative Task Scheduling

A cooperative task scheduler allows a task to run until it can't or doesn't want to run any more. In this way, the tasks must cooperate. A task may have an infinite loop that never completes, and if he doesn't pass or yield control back, will consume 100% of the processor, 100% of the time. This is a sub-optimal situation. Cooperative task scheduling can be improved by language support for asynchronous architectures, where the language can implicitly be aware when a task can no longer make progress. This results in the task being immediately pulled off the active queue, and another ready and waiting task to be selected to do some work. This doesn't solve the tight-loop problem, where a task must still cooperatively pass control back to the kernel.

Drogue Device Architecture

Given the above background, along with implementing an untold number of proofs-of-concept, we've selected the following key points for our system:

  • Actor-based: State is held by an actor, accessed/mutated only by that actor, in response to messages.
  • Cooperative Scheduling: Using Rust's async/await support, actors attempt to be non-blocking and share the processor.
  • Message-Passing: Support notifications, requests/responses, and an event-bus.

The Actors

First, being an actor system, the primary point-of-interest is the actor. Drogue Device supports two flavors of actor: Plain Ol' Actors and Interrupt Actors.

Plain Ol' Actors

The general case of an actor is a component that contains state, and can manipulate it in response to messages. Their only interface with the outside world is messages. To ensure that no other code can directly twiddle an actor, the primary handle on an actor is an Address<A> instance. It is through the Address<A> of an actor A that any part of the system can communicate with the actor. All methods upon Address<A> are immutable, and the addresses can be freely clone()'d and shared around.

As noted above, there are two types of message interactions possible with an actor through it's Address<A>:

Notify

pub fn notify<M>(&self, message: M)
where
    A: NotifyHandler<M> + 'static,
    M: 'static,
{
    ... elided ...
}

This signature demonstrates a few points:

  • The system is ultimately 'static-centric.
  • You can send a message M if the underlying Actor implements the NotifyHandler<M> trait.

A notify(...) is a non-blocking synchronous call, meaning that it won't stop the processing and it can be called from any context, sync or async. It ultimately enqueues the message into a FIFO for the actor. Some systems refer to this FIFO as the actor's "inbox". Messages in the FIFO are indeed processed in a first-in-first-out order. This also means that the message is not processed immediately upon the call of notify(...). This gives this method fire and forget semantics. The actor will be mostly (see NB below) guaranteed to eventually process the notification, at some point in the future.

NB: Currently the FIFO is set to a depth of 16, and if it overflows messages will be silently discarded. We understand this is sub-optimal. The intention is that once const generics are available in stable Rust we will allow per-actor queue depth configuration. Additionally, we are considering how different overflow strategies may be applied per-actor.

On the Actor side, processing is handled through an implementation of the NotifyHandler<M> trait, for each type of message that is considered acceptable:

pub struct On;
pub struct Off;

impl<P,A> NotifyHandler<On> for SimpleLED<P,A>
where
    P: OutputPin,
    A: ActiveOutput,
{
    fn on_notify(&'static mut self, message: On) -> Completion {
        ... elided ...
    }
}

impl<P,A> NotifyHandler<Off> for SimpleLED<P,A>
where
    P: OutputPin,
    A: ActiveOutput,
{
    fn on_notify(&'static mut self, message: Off) -> Completion {
        ... elided ...
    }
}

Request

Sometimes it's useful to have a somewhat synchronous interactino with an actor in an request/response type of cycle. The Address<A> type also provides that capability:

pub async fn request<M>(&self, message: M) -> <A as RequestHandler<M>>::Response
where
    A: RequestHandler<M> + 'static,
    M: 'static,
{
    ... elided ...
}

The first thing to note is that because Address<A> is a type and not a trait, it can indeed support async methods. The second thing to note is that request(...) is an async method. Upon calling this method, a Future is ultimately returned, and the called must .await the response, per usual Rust async semantics.

It is the async nature of this method that allows actors to talk to other actors and remain non-blocking and cooperatively schedulable.

As with the notify(...) method, this method is also considered immutable, and uses the underlying actor's FIFO. The same FIFO is shared between all types of messages that an actor can be notified or requested.

On the Actor side, as with the NotifyHandler<M>, there is a RequestHandler<M> trait to be implemented for each type of request an actor needs to be able to handle.

pub struct Lock;

impl<T> RequestHandler<Lock> for Mutex<T>
where
    T: 'static,
{
    type Response = Exclusive<T>;

    fn on_request(&'static mut self, message: Lock) -> Response<Self::Response> {
        ... elided ...
    }
}

Interrupt Actors

Extending on the Plain Ol' Actors are Interrupt actors, which in addition to doing all the things an actor can do, also is cognizant and connected to hardware interrupts. Interrupts are special, because they can arrive at any time, and they are not initiated as a message from another actor. So interrupts are treated somewhat specially.

An interrupt firing can be considered a special type of message, without content. To that end, the Interrupt trait brings one more method to implement: on_interrupt(...):

impl<D, PIN> Interrupt for Button<D, PIN>
where
    D: Device + EventHandler<ButtonEvent> + 'static,
    PIN: InputPin + ExtiPin,
{
    fn on_interrupt(&mut self) {
        ... elided ...
    }
}

Message-handling and async

Since Drogue Device attempts to provide guardrails for performing asynchronous operations where possible, but Rust doesn't currently support async in traits, you may have noticed that all of the aforementioned NotifyHandler<M> and RequestHandler<M> methods were not, at all, async. Additionally, their return value in the signatures may have appeared slightly odd. Also, each method takes a &'static mut self reference. It's `static because the system is static, and Futures work best with `static`. It's &mut because the scheduler guarantees that the actor's methods are serialized and ensures that a given handler is the only reference (mutable or not) for the actor at any given point in time. This allows the actor to free, without locks, mutate his own state.

on_notify

Let's look at the fire-and-forget NotifyHandler<M> on an actor.

fn on_notify(&'static mut self, message: M) -> Completion;

Even though Address<A>::notify(...) returns nothing, the handler implemented on the actor must return a Completion, which is where the opportunity to perform asynchronous processing arrives.

If the processing is quick and non-blocking, a simple implementation can do whatever work it needs to do and return Completion::Immediate to signal that the message was immediately processed to completion.

An ::immediate() function is provided to create a Completion::Immediate for the return value.

fn on_notify(&'static mut self, message: M) -> Completion {
    self.counter += 1;
    Completion::immediate()
}

If in response to a message an actor needs to perform some other action, such as making an async request to another actor, a Completion::Defer(...) is available, which ultimately wraps an async block.
The executor will attempt to avoid a context-switch and begin executing the returned async block as far as possible before swapping to another task.

A ::defer(...) function is provided to create a Completion::Defer(...) for the return value.

fn on_notify(&'static mut self, message: M) -> Completion {
    Completion::defer( async move {
        self.counter += self.other_actor.request( SomeMessage ).await;
    })
}

on_request

As with on_notify, the on_request(...) method also takes the same flavor of &'static mut self and provides both an immediate and defer variant of Response<T>, but allow for the return-value as specified in the associated type of RequestHandler.

An immediate response is allowed if async is not required:

fn on_request(&'static mut self, message: M) -> Response<Self::Response> {
    Response::immediate(42)
}

If an async block is needed:

fn on_request(&'static mut self, message: M) -> Response<Self::Response> {
    Response::defer( async move {
        let response = self.some_other_actor.rqeuest( AnotherMessage ).await;
        response.favorite_cheese
    } )
}

on_interrupt

For actors that also implement Interrupt, their on_interrupt(...) method may be called when the associated interrupt fires. Since interrupts must return quickly, this is a simple, normal, non-async method without opportunity of deferring to an async block:

fn on_interrupt(&mut self) {
    self.some_other_actor.notify( SomethingHappened );
}

This also implies that Interrupt implementations can only use notify(...) on other actors, and not make asynchronous request(...) calls.

General async Executor

There isn't one.

Drogue Device explicitly does not support arbitrary spawn(...)ing of async blocks. All asynchronous activity takes place through coordination of actors and their FIFO queues.

There is an async executor, you're just not allowed to touch it directly.

Wiring it all together

So far we've explored the various actor, address, message-handlers and async aspects, so let's now examine how we connect it all together.

The Device

At the root of an actor system is a Device. This is a user-constructed type that ultimately ends up holding all of the other Actor and Interrupts, but indirectly. When the system is started, the entire tree, from the Device on down is moved into a 'static context, which is how all of the message-handlers end up taking a &'static mut self.

This example will use an STM32 B-L4S5I-IOT01A board, but the same principals apply to others, even if the pinout is different.

The Type

For ease, a simple type with public members can be defines. Since nothing should actually hold an Actor directly, each Actor or Interrupt is ultimately held in a matching context. Type aliases have been used to simplify the signatures since generics are quite prevalent.

An actor has been specified for each of the two LEDs on this board, along with an interrupt actor for the button. Additionally, two Blinker actors and a timer actor are also specified. Since hardware timers are interrupt-driven, it is an interrupt actor.

pub struct MyDevice {
    pub ld1: ActorContext<Ld1Actor>,
    pub ld2: ActorContext<Ld2Actor>,
    pub blinker1: ActorContext<Blinker1Actor>,
    pub blinker2: ActorContext<Blinker2Actor>,
    pub button: InterruptContext<ButtonInterrupt>,
    pub timer: InterruptContext<TimerActor>,
}

The Device Implementation

A single method is required when implementing Device: mount(...).

When the system is started, the device will be automatically mounted into it. Since the device holds the other actors and interrupts, it must propagate the mount down to the children in order to activate them within the system also. Upon mounting each child, its Address<A> will be returned.

Each Actor can optionally implement a Bind<OTHER_ACTOR:Actor> method to have a way to inject another actor's Address<A> into it. This is one way inform an actor of another actor during the mount(...) cycle.

impl Device for MyDevice {
    fn mount(
        &'static mut self,
        bus_address: &Address<EventBus<Self>>,
        supervisor: &mut Supervisor,
    ) {
        let ld1_addr = self.ld1.mount(supervisor);
        let ld2_addr = self.ld2.mount(supervisor);

        let blinker1_addr = self.blinker1.mount(supervisor);
        let blinker2_addr = self.blinker2.mount(supervisor);

        let timer_addr = self.timer.mount(supervisor);

        blinker1_addr.bind(&timer_addr);
        blinker1_addr.bind(&ld1_addr);

        blinker2_addr.bind(&timer_addr);
        blinker2_addr.bind(&ld2_addr);

        let button_addr = self.button.mount(supervisor);
        button_addr.bind(bus_address);
    }
}

In the code above, each Blinker actor gets bind(...) called twice: once for the LED it needs to blink, and once for the shared Timer.

The Button actor gets bound to the address of the EventBus which we will touch on later.

The Setup

Using bare-metal Rust, we set up the actors and the device in a normal #[entry]:

#[entry]
fn main() -> ! {
    rtt_init_print!();
    log::set_logger(&LOGGER).unwrap();
    log::set_max_level(log::LevelFilter::Debug);

    let mut device = Peripherals::take().unwrap();

    log::info!("[main] Initializing");
    let mut flash = device.FLASH.constrain();
    let mut rcc = device.RCC.constrain();
    let mut pwr = device.PWR.constrain(&mut rcc.apb1r1);
    let clocks = rcc
        .cfgr
        .sysclk(80.mhz())
        .pclk1(80.mhz())
        .pclk2(80.mhz())
        .freeze(&mut flash.acr, &mut pwr);

    let mut gpioa = device.GPIOA.split(&mut rcc.ahb2);
    let mut gpiob = device.GPIOB.split(&mut rcc.ahb2);
    let mut gpioc = device.GPIOC.split(&mut rcc.ahb2);
    let mut gpiod = device.GPIOD.split(&mut rcc.ahb2);

    // == LEDs ==

    let ld1 = gpioa
        .pa5
        .into_push_pull_output(&mut gpioa.moder, &mut gpioa.otyper);

    let ld1 = SimpleLED::new(ld1, Active::High);

    let ld2 = gpiob
        .pb14
        .into_push_pull_output(&mut gpiob.moder, &mut gpiob.otyper);

    let ld2 = SimpleLED::new(ld2, Active::High);

    // == Blinker ==

    let blinker1 = Blinker::new(Milliseconds(500u32));
    let blinker2 = Blinker::new(Milliseconds(1000u32));

    // == Button ==

    let mut button = gpioc
        .pc13
        .into_pull_up_input(&mut gpioc.moder, &mut gpioc.pupdr);

    button.make_interrupt_source(&mut device.SYSCFG, &mut rcc.apb2);
    button.enable_interrupt(&mut device.EXTI);
    button.trigger_on_edge(&mut device.EXTI, Edge::RISING_FALLING);

    let button = Button::new(button, Active::Low);

    // == Timer ==

    let mcu_timer = McuTimer::tim15(device.TIM15, clocks, &mut rcc.apb2);
    let timer = Timer::new(mcu_timer);

    // == Device ==

    let device = MyDevice {
        ld1: ActorContext::new(ld1).with_name("ld1"),
        ld2: ActorContext::new(ld2).with_name("ld2"),
        blinker1: ActorContext::new(blinker1).with_name("blinker1"),
        blinker2: ActorContext::new(blinker2).with_name("blinker2"),
        button: InterruptContext::new(button, EXTI15_10).with_name("button"),
        timer: InterruptContext::new(timer, TIM15).with_name("timer"),
    };

    device!( MyDevice = device; 1024 );
}

Each actor is directly constructed and configured within the #[entry], wrapped in the appropriate context and attached to the device. You will notice that InterruptContext::new(...) also takes an IRQ to be connected to the underlying actor. When the associate interrupt fires, the actor's on_interrupt(...) will be triggered.

Optionally, each context can be associated with a name, which can make debugging or logging easier in a multi-actor system, particular when working with several actors of the same type.

The device!(...) magic

The last line of the #[entry] specifies the type implementing Device, the device itself, plus the size of memory to reserve for asynchronous book-keeping. In this case, 1024 bytes is allotted.

The device!(...) macro never returns and is a suitable last line for #[entry].

At this point, the system is running.

Explore the Example

All of the example actors in our system are part of the drogue-device crate.

LEDs

Each LED is a SimpleLED actor, which is parameterized by the appropriate digital OutputPin and an Active state which describes if it's active-high or active-low. Each LED can handle a notification of simple On or Off messages:

impl<P,A> NotifyHandler<On> for SimpleLED<P,A>
where
    P: OutputPin,
    A: ActiveOutput,
{
    fn on_notify(&'static mut self, message: On) -> Completion {
        self.turn_on();
        Completion::immediate()
    }
}

impl<P,A> NotifyHandler<Off> for SimpleLED<P,A>
where
    P: OutputPin,
    A: ActiveOutput,
{
    fn on_notify(&'static mut self, message: Off) -> Completion {
        Completion::defer(async move {
            self.turn_off();
        })
    }
}

Magic with Address<SimpleLED<...>>

Since the SimpleLED is implemented within the drogue-device crate, along with Address, we can actually extend Address<SimpleLED<...>> to provide handy API methods.

impl<S> Address<S>
where
    S: NotifyHandler<Off> + NotifyHandler<On>,
    S: Actor + 'static,
{
    pub fn turn_on(&self) {
        self.notify(On);
    }

    pub fn turn_off(&self) {
        self.notify(Off);
    }
}

Now, anyone holding an Address<SimpleLED<...>> does not even have to think in terms of notify(On) or notify(Off), but can directly call turn_on() and turn_off().

Timer

The Timer actor was constructed using a hardware timer of our platform, and wired to the appropriate interrupt. The timer is a shared resource, capable of providing asynchronous delays or scheduling future delivery of messages.

Since an asychronous delay is asynchronous and must be awaited, it is implemented as a RequestHandler<Delay<...>>.

impl<T: HalTimer, DUR: Duration + Into<Milliseconds>> RequestHandler<Delay<DUR>> for Timer<T> {
    type Response = ();

    fn on_request(&'static mut self, message: Delay<DUR>) -> Response<Self::Response> {
        ... elided ...
    }
}

The request has no meaningful response type, but the .await from the caller will only return once the delay has expired.

We'll be using the Schedule message type, which simply asks the timer to deliver, via notify(...) some message to some address at some point in the future.

impl<T, E, A, DUR> NotifyHandler<Schedule<A, DUR, E>> for Timer<T>
where
    T: HalTimer,
    E: Clone + 'static,
    A: Actor + NotifyHandler<E> + 'static,
    DUR: Duration + Into<Milliseconds> + 'static,
{
    fn on_notify(&'static mut self, message: Schedule<A, DUR, E>) -> Completion {
        ... elided ...
    }
} 

Blinker

If you recall, our Blinker was bound to both the LED and the timer. Since its own address, along with that of the LED and Timer are not known at construction time, it has Option<Address<...>> fields for populating later, during start-up or via bind(...) calls.

pub struct Blinker<S, T>
where
    S: Switchable,
    T: HalTimer,
{
    led: Option<Address<S>>,
    timer: Option<Address<Timer<T>>>,
    address: Option<Address<Self>>,
    delay: Milliseconds,
}

The Blinker implements some methods on the Actor trait which allow it to learn of its own Address and to perform some action at start-up. In this case, it schedules a message of State::On back to itself after delay milliseconds when it starts.

impl<S, T> Actor for Blinker<S, T>
where
    S: Switchable,
    T: HalTimer,
{
    fn mount(&mut self, address: Address<Self>)
    where
        Self: Sized,
    {
        self.address.replace(address);
    }

    fn start(&'static mut self) -> Completion {
        self.timer.as_ref().unwrap().schedule(
            self.delay,
            State::On,
            self.address.as_ref().unwrap().clone(),
        );
        Completion::immediate()
    }
}

When the timer delivers that future-scheduled message, it will take action on it, and scheduler another message to itself to get a blinking LED, toggling forever.

impl<S, T> NotifyHandler<State> for Blinker<S, T>
where
    S: Switchable,
    T: HalTimer,
{
    fn on_notify(&'static mut self, message: State) -> Completion {
        match message {
            State::On => {
                self.led.as_ref().unwrap().turn_on();
                self.timer.as_ref().unwrap().schedule(
                    self.delay,
                    State::Off,
                    self.address.as_ref().unwrap().clone(),
                );
            }
            State::Off => {
                self.led.as_ref().unwrap().turn_off();
                self.timer.as_ref().unwrap().schedule(
                    self.delay,
                    State::On,
                    self.address.as_ref().unwrap().clone(),
                );
            }
        }
        Completion::immediate()
    }
}

We also can envision wanting to change the speed of the blinking during the course of the system running, so we provide a NotifyHandler<AdjustDelay> to the Blinker:

pub struct AdjustDelay(Milliseconds);

impl<S, T> NotifyHandler<AdjustDelay> for Blinker<S, T>
where
    S: Switchable,
    T: HalTimer,
{
    fn on_notify(&'static mut self, message: AdjustDelay) -> Completion {
        self.delay = message.0;
        Completion::immediate()
    }
}

Once again, also, we add some fluent API to instances of Address<Blinker<...>:

impl<S, T> Address<Blinker<S, T>>
where
    Self: 'static,
    S: Switchable,
    T: HalTimer,
{
    pub fn adjust_delay(&self, delay: Milliseconds) {
        self.notify(AdjustDelay(delay))
    }
}

Button (and the EventBus)

The Button is an interrupt-based actor. While we could have used bind(...) to target a specific recipient of its ButtonEvent::Pressed or ButtonEvent::Released messages, that would have tightly-bound the target to the semantics of the button.

Instead, the Button is bound to the Address<EventBus<...>> during the Device mount, in order to delegate what action to take to a loosely-coupled actor.

pub struct Button<D: Device, PIN> {
    pin: PIN,
    active: Active,
    bus: Option<Address<EventBus<D>>>,
}

impl<D, PIN> Bind<EventBus<D>> for Button<D, PIN>
where
    D: Device,
{
    fn on_bind(&'static mut self, address: Address<EventBus<D>>) {
        self.bus.replace(address);
    }
}

When the interrupt fires, its on_interrupt(...) method is invoked, where it calls notify(...) on the EventBus<...> address.

impl<D, PIN> Interrupt for Button<D, PIN>
where
    D: Device + EventHandler<ButtonEvent> + 'static,
    PIN: InputPin + ExtiPin,
{
    fn on_interrupt(&mut self) {
        if self.pin.check_interrupt() {
            match self.active {
                Active::High => {
                    if self.pin.is_high().ok().unwrap() {
                        self.bus
                            .as_ref()
                            .unwrap()
                            .publish(ButtonEvent::Pressed);
                    } else {
                        self.bus
                            .as_ref()
                            .unwrap()
                            .publish(ButtonEvent::Released);
                    }
                }
                Active::Low => {
                    if self.pin.is_low().ok().unwrap() {
                        self.bus
                            .as_ref()
                            .unwrap()
                            .publish(ButtonEvent::Pressed);
                    } else {
                        self.bus
                            .as_ref()
                            .unwrap()
                            .publish(ButtonEvent::Released);
                    }
                }
            }
            self.pin.clear_interrupt_pending_bit();
        }
    }
}

What is the EventBus?

So far we've glossed over the EventBus. The EventBus is your device. Your device provides all logic for what happens to messages sent to the EventBus address, in order to apply application semantics.

Just as actors and interrupts can implement NotifyHandler<M> and RequestHandler<M>, your Device can (and must) implement EventHandler<M> for all events destined to the EventBus.

impl EventHandler<ButtonEvent> for MyDevice {
    fn on_event(&'static mut self, message: ButtonEvent)
    where
        Self: Sized,
    {
        match message {
            ButtonEvent::Pressed => {
                log::info!("[{}] button pressed", ActorInfo::name());
                self.blinker1.address().adjust_delay(Milliseconds(100u32));
            }
            ButtonEvent::Released => {
                log::info!("[{}] button released", ActorInfo::name());
                self.blinker1.address().adjust_delay(Milliseconds(500u32));
            }
        }
    }
}

Since your Device maintains the contexts for each component, it can obtain the address for each, and send notifications down the line. In this case, when the button is pressed, our Device translates that into adjusting the delay for the Blinker that controls LED 1. Upon releasing the button, it re-adjusts the delay.

The end result is pressing the button speeds up the blinking until the button is released.

And the button is ignorant of the LED, and the LED is ignorant of the button. You device provides the glue between two disparate component message semantics.

Other Bits in the Tin

While LEDs and Buttons are nice, drogue-device also provides within the crate concurrency actors, such as Mutext<T> and Semaphore. Additionally, Ulf has recently worked on UART support that does not require 'static messages.

We also aim to include device drivers directly in-tree.

Future directions

There's a lot of work left to do, much of which will be bringing significantly more async support to various MCUs and HALs. There is probably a lot of API clean-up that can be done, in order to make implementing actors easier. We mentioned way way earlier, we intend to make each actor's FIFO more configurable to allow for more tailored overflow behaviour. Being able to dynamically adjust the priority of the async tasks is also a goal.

We welcome any and all contributions to the project. Feel free to drop by our Matrix chat (see sidebar) or use our Discourse server.

Where to find it?

Currently we've not published a crate to crates.io because too much is in flux. But you can find it on our GitHub: