UdpSocket not properly closed after session termination

[emliunix]

It's guaranteed that UdpSocket will release the resources when dropped. Eg. FileDesc::Drop() on unix platforms.

In my case, however, it's not dropped. I guess it's still kept somewhere in the async engine.

async fn handle_session() {
    let udp_port = tokio::net::UdpSocket::bind("...").await?;
    join!(other_task, handle_udp(udp_port))?;
}

how UdpSocket guaranteed to close on unix

The hierarchy is like:

• std::net::UdpSocket
• std::sys_common::UdpSocket
• std::sys::unix::net::Socket
• std::sys::unix::fd::FileDesc::Drop
• libc::close

[emliunix]

I think that's simply because the old session hasn't terminated in the first place. The session blocks at receiving next TCP message while the remote already exited.

And my CPU fan indicates the blocking must be some kind of loop.

[emliunix]

Seems it's still related to UdpSocket. After first UDP error, which is udp recv error, the CPU goes up to 100%. Inspect with a debugger shows it's stalled in UdpSocket::recv.

So the weired part is:

• The error message shows only once
• So some loop in the recv call.
• And it should be UdpSocket::send that's the root cause. Maybe the socket (fd) is corrupted.

Maybe I can write a more clean code with only the UDP part to reproduce this issue.

And maybe I can stick to the principle that UDP receive error is unrecoverable.

[emliunix]

Sticking to the principle makes the program running normally. At least no CPU 100% in low-level code.

And now I need to figure out ways to manage my UDP connection. The problems are:

• The unreliability of UDP should be isolated to that UdpSocket, if the remote is not listening, the packet is OK to be discarded, but the program itself should not crash. Lessons learnt above is that when send failed, the recv part exited too, the udpSocket cannot be reused. So maybe I need to recreate UdpSocket on failure. And the session should be aware of readiness of UdpSocket. Also, it's a concern whether frequent recreation of UdpSocket is a problem.
• When TCP part (Session) is requested to terminate, the UdpSocket should be terminated as well. And in my simple test case. I need to figure out a way to interrupt UdpSocket::recv. Otherwise it's waiting indefinitely and the session hence never terminates.