Soundness issues

[Dirbaio]

• Wakers contain a pointer to stack-allocated tasks, and nothing prevents the task from being deallocated while there still are wakers pointing to it.
• Stack-allocated tasks are not pinned. You can mem::forget the JoinHandle (preventing the task from getting unlinked by JoinHandle drop) which releases the borrow on the task, and then move the task, which will break the linked list pointers.
• cargo miri test and cargo miri run --example ... fail. I haven't looked deep in the code to know whether this is due to the above issues, or to other issues.

[sammhicks]

As of 9ff2b37f559a99f84c484edc6c9d258b43119fa6, Miri no longer complains.

[Dirbaio]

the other issues still stand though!

it's fundamentally unsound to allocate tasks in the stack and make wakers point directly to the tasks.

[sammhicks]

@Dirbaio I think I have a solution, currently sitting in pinned_tasks. Both futures_util::pin_mut! and core::pin::pin! take ownership of the task, ensuring that they cannot be core::mem::forgeted.

[Dirbaio]

That doesn't fix the 1st issue "Wakers contain a pointer to stack-allocated tasks, and nothing prevents the task from being deallocated while there still are wakers pointing to it.".

Wakers are 'static. Once you get a waker pointing to a task, you can store it literally anywhere, forever. If the task is pinned then you're guaranteed its Drop will run, but there's no way to make Drop "recall" all the wakers pointing to the task that could be stored anywhere across the whole program.

It's possible to do if you add indirection to the wakers. Like, store a "task ID" in wakers, executor contains an id->pointer map so you can mark the id as "stale" on task drop to "deactivate" all the wakers. This adds complexity and hurts perf tho.

I've looked at this hard over the years and my conclusion is the only two good designs are either to alloc+refcount tasks (what most std executors do) or statically allocate tasks so they live forever (what embassy-executor and RTIC do). There's no way to stack-allocate tasks with good perf.

[sammhicks]

Yes, Wakers are the unsolvable issue.

I'd originally written this crate way back in 2020 when as far as I could see there wasn't really a way of doing async on bare metal. Since then both embassy and RTIC have flourished, and if fact I use embassy for my personal projects. As such, I think it's time to deprecate and retire nostd_async.

Do you by any chance know how I might do this?

Possibilities include:

• Posting on the Github page that the project is retired
• Yanking all versions of nostd_async
• Publishing a new version with a compile error notifying users to use something else