resource leak in nested task

[tabokie]

See reproducible case below (in the context of TiKV's future_pool.rs), an entire ThreadPool is leaked which I presume will be run in detached mode.

    #[test]
    fn test_shutdown_yatp_pool() {
        let outer = Builder::new(DefaultTicker {})
            .thread_count(1, 1, 1)
            .build_future_pool();
        let inner = Builder::new(DefaultTicker {})
            .thread_count(1, 1, 1)
            .build_future_pool();

        let inner_clone = inner.clone();
        outer.spawn(async move {
            let (tx, rx) = oneshot::channel();
            inner_clone.spawn(async move {
                println!("before bed");
                std::thread::sleep(std::time::Duration::from_secs(100));
                println!("slept over");
                tx.send(()).unwrap();
            }).unwrap();
            rx.await.unwrap();
        }).unwrap();

        std::thread::sleep(std::time::Duration::from_secs(1));

        drop(outer);

        // assertion fails with 2.
        assert_eq!(Arc::strong_count(&inner.inner), 1);
    }

[tabokie]

cc @BusyJay @sticnarf for help, I'm not familiar with async Rust to understand this.

[sticnarf]

I would say this is expected.

Let's call this part fut:

        async move {
            let (tx, rx) = oneshot::channel();
            inner_clone.spawn(async move {
                println!("before bed");
                std::thread::sleep(std::time::Duration::from_secs(100));
                println!("slept over");
                tx.send(()).unwrap();
            }).unwrap();
            rx.await.unwrap();
        }

fut captures inner_clone by move, so inner_clone won't drop until fut drops.

When polling rx in fut, an Arc of fut (not really, but equivalent) will be included in the Waker. The Waker is passed to the internal of the channel, so the future can be woken up when tx sends something.

So, it's the oneshot channel that holds a reference count of fut, and fut holds inner_pool.

[tabokie]

Interesting. Thanks for explaining. I'll just work around this by manually closing the inner before shutdown.