Features `time` and `executor-agnostic` don't work together

[diondokter]

trafficstars

Timing data is stored in the task header, so in the function register_timer it tries to get the task header from the waker.

But this waker is not made by embassy and a panic is generated.

Is this a bug? Is this intentional or a "won't fix"? (If so, maybe it shouldn't compile or at least that should be documented)

What I'm running into is that I have an async library that can be used in embassy. It needs some timer things. Then I also want to write some tests using futures-test which then get the aforementioned panic.

[diondokter]

I can kinda do it if I make an embassy executor:

    #[test]
    fn main() {
        let res = std::panic::catch_unwind(|| {
            static EXECUTOR: Forever<Executor> = Forever::new();
            let executor = EXECUTOR.put(Executor::new());
            executor.run(|spawner| {
                spawner.spawn(main_loop()).unwrap();
            });
        });

        match res {
            Ok(_) => {}
            Err(e) => match e.downcast::<&str>() {
                Ok(message) if *message == "Done" => {},
                Ok(e) => std::panic::resume_unwind(e),
                Err(e) => std::panic::resume_unwind(e),
            },
        }
    }

But the executor will still run even if all tasks have returned, so the only way to stop the test is with a panic that you then catch which is not optimal...

[lulf]

Fyi, for a similar use case of running embassy in a test, I made this: https://github.com/drogue-iot/ector/blob/main/ector/src/testutil.rs#L269-L361

in addition I created a custom macro so you could run a test being handed a context that would signal the runner to stop on drop.

If ppl are interested I could create some 'embassy-test' crate that would allow writing unit tests the same way, but I think ideally making this embassy-agnostic in the tests would be even better so one could reuse existing test frameworks from tokio etc.

[Dirbaio]

It is intentional that the current Timer implementation is tied to the executor. It keeps track of expirations per task, not per timer, which allows it to reuse the task queue as a timer queue, which means it's faster/smaller etc.

It's not "wontfix", I do believe in interoperability so I think it'd be great to get it working on any executor. It'd require a new from-scratch implementation, with a real "timer queue" where creating a Timer adds an entry, dropping it removes it, and something (perhaps a time driver alarm?) that watches for expirations and wakes the wakers. It will be slower/bigger than the current one, and more annoying to use (you'll have to specify the size of the timer queue upfront) so it should be opt-in, enabled by the feature executor-agnostic (just like the current AtomicWaker/WakerRegistration).

I won't work on it, but will happily mentor+review anyone who wants to try!

[Dirbaio]

embassy-time is now executor-agnostic by default. You can enable an executor agnostic queue with the embassy-time/generic-queue feature, or you can enable the embassy-executor-integrated one with embassy-executor/integrated-timers.