Building an Embedded Futures Executor II

27 Jan 2019

Part 2 of my embedded executor journey!

Part 1 ended up getting a little long, so I cut it short while still missing some key features like more flexible task spawning and putting the executor to sleep when there's nothing needing to be polled. This time, we'll fill in those gaps!


As of my last post, our executor was capable of spawning some arbitrary number of tasks via its spawn method and then run them all to completion. This, unfortunately, is pretty limiting. Tasks can only be spawned before the executor is running, and you have to wait until it's done to spawn anything else.

What would be really nice is a way to spawn new tasks while the executor is running. So what needs to happen to make this a thing? If you recall from before, spawning a task involves 1. Adding it to the task registry and 2. Inserting its ID into the queue to be polled. A somewhat naiive approach to this would be to wrap the registry in an Arc<Mutex<_>> so that it can be shared along with the queue. But we already have one lock - it would be nice to not have to add another or to wrap our entire executor in one big lock. So why not simply re-use the queue that we already have in place? Rather than just a queue of task IDs, the queue will also hold new futures to spawn.

enum QueueItem<'a> {
    Spawn(FutureObj<'a, ()>),

// Some aliases for convenience
type Queue<'a> = VecDeque<QueueItem<'a>>;
type QueueHandle<'a, R> = Arc<Mutex<R, Queue<'a>>>;

The interior loop in our run method will now look something like this:

while let Some(item) = self.dequeue() {
    match item {
        QueueItem::Poll(id) => { /* same thing as before */ },
        // You can always go from `FutureObj` -> `LocalFutureObj`
        QueueItem::Spawn(future) => self.spawn(future.into()),

And all we need now is a nice way to package up the queue and present a nice spawning API:

struct Spawner<'a, R>(QueueHandle<'a, R>);

impl<'a, R> Spawner<'a, R> {
    fn new(handle: QueueHandle<'a, R>) -> Self {

    fn spawn(&mut self, future: FutureObj<'a, ()>) {

With that, we now have a cloneable, Send-able way to spawn new tasks!


The last piece of the puzzle to accomplish all of our originally stated goals is a way to abstract over "sleeping" the executor thread while waiting to be notified of tasks needing to be polled.

My (probably naiive and insufficient) approach to this was to define a simple Sleep trait:

trait Sleep: Default + Clone + Send + Sync + 'static {
    fn sleep(&self);

    fn wake(&self);

Which can then be threaded through our executor down to the QueueWaker:

struct Executor<'a, R, S: Sleep> {
    sleeper: S,

struct QueueWaker<'a, R, S: Sleep> {
    sleeper: S,

In the event loop:

// Before polling the task:
let waker =
        QueueWaker(self.queue.clone(), id, self.sleeper.clone())

// After the "registry empty" check:

In the Wake::wake method:


While this may not be sufficient for every use case, it seems to work well enough for now. It could be implmented as an AtomicBool flag that's checked in a core::sync::atomic::spin_loop_hint loop, or something fancier like a Condvar. On Cortex-M systems, you can use the wfi or wfe instructions, assuming events are going to come in as interrupts or events from other cores.


Aaaand that's pretty much it! There's a bit more to the final (for now) implementation to make things more ergonomic or efficient, and plenty of R: RawMutex + Sync + Send bounds that I've glossed over to cut down on the noise, but the core of my embedded executor is all there. You can find it on and GitLab if you want to dig deeper into its internals and documentation.