TcpStream::connect_timeout doesn't fail to connect to a nonexistent server on Linux

[sfackler]

use std::net::TcpStream;
use std::time::Duration;

fn main() {
    let sock = "127.0.0.1:1234".parse().unwrap();
    let timeout = Duration::from_secs(3);
    TcpStream::connect_timeout(&sock, timeout).unwrap();
}

This program succeeds on Linux, even though it shouldn't (assuming you don't have a listener on 1234). Trying to do anything with the socket produces Error { repr: Os { code: 107, message: "Transport endpoint is not connected" } }.

[sfackler]

Interestingly it does behave as expected on macOS.

[karizon]

Works under WSL as well. Native linux it fails spectacularly.

[sfackler]

From an strace, it looks like Linux sets all of POLLOUT, POLLERR, and POLLHUP at once. We treat POLLOUT as success, but it seems like we should be checking for POLLERR instead:

socket(AF_INET, SOCK_STREAM|SOCK_CLOEXEC, IPPROTO_IP) = 3
ioctl(3, FIONBIO, [1])                  = 0
connect(3, {sa_family=AF_INET, sin_port=htons(80), sin_addr=inet_addr("127.0.0.1")}, 16) = -1 EINPROGRESS (Operation now in progress)
ioctl(3, FIONBIO, [0])                  = 0
clock_gettime(CLOCK_MONOTONIC, {tv_sec=632431, tv_nsec=149012221}) = 0
clock_gettime(CLOCK_MONOTONIC, {tv_sec=632431, tv_nsec=149240049}) = 0
poll([{fd=3, events=POLLOUT}], 1, 2999) = 1 ([{fd=3, revents=POLLOUT|POLLERR|POLLHUP}])

[sfackler]

Fix here: #45269