windows: `TcpStream::shutdown` does not wake up blocking reads

[lukas-code]

Description

The docs for TcpStream::shutdown currently guarantee that a shutdown with Shutdown::Read must wake up pending operations immediately:

This function will cause all pending and future I/O on the specified portions to return immediately with an appropriate value (see the documentation of Shutdown).

The docs for Shutdown::Read explicitly says that currently blocked reads must return:

All currently blocked and future reads will return Ok(0).

However, that is currently not the case on *-pc-windows-* platforms. Instead, a currently blocked read will stay blocked forever when shutdown is called with Shutdown::Read in a different thread. The same also happens with Shutdown::Both.

This is causing the close_read_wakes_up test to fail spuriously due to a race condition, for example in #121523 (comment) and #120543 (comment). The test only passes if the shutdown happens before the read.

Repro

I tried this code:

use std::io::Read;
use std::net::{Ipv6Addr, Shutdown, TcpListener, TcpStream};
use std::thread;
use std::time::Duration;

fn main() {
    let listener = TcpListener::bind((Ipv6Addr::LOCALHOST, 0)).unwrap();
    let listener_addr = listener.local_addr().unwrap();
    let mut serverbound_stream = None;
    thread::scope(|scope| {
        // 1. `accept` and `connect` concurrently
        let _clientbound_stream = scope.spawn(|| {
            let (clientbound_stream, _) = listener.accept().unwrap();
            clientbound_stream
        });
        let serverbound_stream =
            serverbound_stream.insert(TcpStream::connect(listener_addr).unwrap());

        // 3. shutdown read during blocking read
        scope.spawn(|| {
            thread::sleep(Duration::from_secs(1)); // just to be sure
            serverbound_stream.shutdown(Shutdown::Read).unwrap();
        });

        // 2. blocking read
        let count = (&*serverbound_stream).read(&mut [0]).unwrap();
        assert_eq!(count, 0);
    });
}

I expected to see this happen: The .shutdown(Shutdown::Read) makes the blocking read return immediately with Ok(0).

Instead, this happened: The blocking read does not return, making the program hang indefinitely.

Meta

rustc --version --verbose:

rustc 1.78.0-nightly (381d69953 2024-02-24)
binary: rustc
commit-hash: 381d69953bb7c3390cec0fee200f24529cb6320f
commit-date: 2024-02-24
host: x86_64-unknown-linux-gnu
release: 1.78.0-nightly
LLVM version: 18.1.0

Tested on these platforms:

• x86_64-pc-windows-msvc with Microsoft Windows Server 2019 (GitHub CI)
• x86_64-pc-windows-gnu with wine 9.1 on Linux

Related

• std: Ignore close_read_wakes_up on Windows #38867 disabled the test due to "We don't document this as a guarantee for shutdown", except that we do and also already did when that PR was opened.
• Enable close_read_wakes_up test on Windows #94714 enabled the test again as "works on my machine".

[lukas-code]

I think there is no (simple) way to make the implementation conform to the docs, so we should probably just remove the guarantees about currently blocked operations from the docs and only talk about future operations.

Note that POSIX shutdown also only talks about subsequent operations and not currently blocked ones.

@rustbot label A-docs T-libs-api O-windows A-io -needs-triage