[Help wanted] Weird behavior that Tokio went asleep forever in certain situations

[cxw620]

Version

1.45.0 ~ 1.46.1

Platform

Linux HANTONG-DEV 6.15.4-x64v3-xanmod1 #0~20250627.g1f365b0 SMP PREEMPT_DYNAMIC Fri Jun 27 15:24:34 UTC x86_64 GNU/Linux

Description

Imagine a scenario of performing bidirectional copying between two TcpStreams using tokio.

Inside poll_copy:

    pub fn poll_copy(
        &mut self,
        cx: &mut Context<'_>,
        mut r: Pin<&mut R>,
        mut w: Pin<&mut W>,
    ) -> Poll<io::Result<()>> {
        loop {
            match self.state {
                TransferState::Running => {
                    while !self.pipe.closed() {
                        crate::trace!(ctx = ?self, "loop `poll_copy`");

                        let _ = self.poll_flush(cx, w.as_mut(), false)?;
                        ready!(self.poll_splice_drain(cx, r.as_mut()))?;
                        ready!(self.poll_splice_pump(cx, w.as_mut()))?;
                    }

                    self.state = TransferState::ShuttingDown;
                }
                TransferState::ShuttingDown => {
                    // ! Must shutdown, or Tokio will hang forever.
                    ready!(w.as_mut().poll_shutdown(cx))?;

                    self.state = TransferState::Finished;
                }
                TransferState::Finished => break Poll::Ready(Ok(())),
            }
        }
    }

The example is available here: https://github.com/hanyu-dev/tokio-splice2, please comment out https://github.com/hanyu-dev/tokio-splice2/blob/b7fcec03bbd53c5329c2727bfa8e86e5b8c0e2a6/src/splice.rs#L574-L575 (or change to poll_flush).
Just run cargo run --release --example proxy, start a iperf3 server iperf3 -s -p 5202 and do some speed tests with iperf3 -c 127.0.0.1 -Z -t 5 -p 5201 -R (must be -R).

If I change poll_shutdown to poll_flush, Tokio would go asleep forever sporadically (if you are lucky enough, the first speed test attempt would immediately trigger this) and iperf3 would receive nothing, tokio-console said that both two tasks (copy a -> b, copy b -> a) are parked and idle, one had no unused waker, while the other one had two unused wakers.

When I force terminate iperf3, Tokio could be woken up normally for the I/O Error.

Maybe I did something wrong and I'm looking forward to corrections or suggestions or tips.