Possible deadlock when a waker is replaced

[satakuma]

Version
1.25.0

Description

There are some types in tokio::sync which keep a shared state behind a mutex and expose methods to modify it through a shared reference. For example: Notify::notify_waiters(&self), Semaphore::add_permits(&self, n: usize).
It is possible to trigger a deadlock with custom wakers backed by ArcWake, which call those methods in their destructors. This can happen because tokio drops a waker while holding the lock to the shared state, and the waker tries to re-enter the lock in its Drop impl. Examples of wakers dropped with the lock held: Notify::notified, Semaphore::acquire.

To trigger a deadlock, you can poll a future twice with two different wakers. At the second poll, the old waker will be replaced and instantly dropped. Examples of deadlocking code:

Notify::notified

Adapted from this test

fn notify_in_drop() {
    use futures::task::ArcWake;
    use std::future::Future;
    use std::sync::Arc;

    let notify = Arc::new(Notify::new());

    struct NotifyOnDrop(Arc<Notify>);

    impl ArcWake for NotifyOnDrop {
        fn wake_by_ref(_arc_self: &Arc<Self>) {}
    }

    impl Drop for NotifyOnDrop {
        fn drop(&mut self) {
            self.0.notify_waiters();
        }
    }

    let mut fut = Box::pin(async {
        notify.notified().await;
    });

    // Second iteration deadlocks.
    for _ in 0..2 {
        let waker = futures::task::waker(Arc::new(NotifyOnDrop(notify.clone())));
        let mut cx = std::task::Context::from_waker(&waker);
        assert!(fut.as_mut().poll(&mut cx).is_pending());
    }
}

Semaphore::acquire

fn release_permits_at_drop() {
    use futures::task::ArcWake;
    use std::future::Future;
    use std::sync::Arc;

    let sem = Arc::new(Semaphore::new(1));

    struct ReleaseOnDrop(Option<OwnedSemaphorePermit>);

    impl ArcWake for ReleaseOnDrop {
        fn wake_by_ref(_arc_self: &Arc<Self>) {}
    }

    let mut fut = Box::pin(async {
        let _permit = sem.acquire().await.unwrap();
    });

    // Second iteration deadlocks.
    for _ in 0..2 {
        let waker = futures::task::waker(Arc::new(ReleaseOnDrop(
            sem.clone().try_acquire_owned().ok(),
        )));
        let mut cx = std::task::Context::from_waker(&waker);
        assert!(fut.as_mut().poll(&mut cx).is_pending());
    }
}

This issue has been partially addressed in the past (#401, PR discussion), but only about calling waker.wake() while holding a lock, not dropping a waker in general.

Simple solution
When an old waker has to be replaced, it can be stored on the side and dropped outside the critical section.

[amab8901]

I'll give it a try

[amab8901]

I'm thinking that a good solution might be the following:

Trigger:
Process P is holding a resource X and requesting another resource Y. 1 second after the request, it still doesn't have the requested resource Y.

Response:
Process A starts a timer and lets go of the held resource X. When the timer hits 5 seconds, it will request the resource X and resource Y and try to execute to completion. If it misses some needed resource and doesn't get it after 1 second, then it repeats the above process.

What are your thoughts on it? Would this be a good approach?

[satakuma]

What you've proposed seems like a way to modify the examples of deadlocking code from the issue description. These examples are only meant to demonstrate the problem which lies within Tokio's internal code. I would like to address this issue by changing the code in Tokio so that such deadlocks are impossible to trigger from a program using Tokio as a library.

To resolve this issue, I would suggest you the following approach:

1. Browse through Tokio's code, starting with the places I've linked in the description (Notify, BatchSemaphore), and look for statements which replace a waker while holding a lock.
2. Modify them so that the old waker is kept in a separate variable, which goes out of scope after the mutex is released.
3. Ensure that it is impossible to produce deadlocking examples such as those from the description.

[amab8901]

Why are we asserting that the future has to be pending, in the Notify test?

assert!(fut.as_mut().poll(&mut cx).is_pending());

[satakuma]

That assert is quite arbitrary, I am uncertain whether it is correct if the code does not deadlock. It was just quicker for me to write the examples that way. Surely you can ignore the result from poll in the second iteration.

[Darksonn]

Closing a semaphore is another instance of this.

[satakuma]

Reopening as it seems this issue got closed on accident (there is still the problem with closing a BatchSemaphore).

[amab8901]

Reopening as it seems this issue got closed on accident (there is still the problem with closing a BatchSemaphore).

What do you mean with BatchSemaphore? I couldn't find any such item in this project

[satakuma]

I mean the Semaphore type from src/sync/batch_semaphore.rs. Sorry for the confusion.

[bbmena]

Should this issue be closed? It looks like BatchSemaphore has been fixed as well.

[Darksonn]

Yes, thank you for pointing that out.