unistd::dup3 never uses the dup3 syscall (even if it might be available)

[tbu-]

nix/src/unistd.rs

Lines 381 to 397 in 70cce1a

	#[inline]
	fn dup3_polyfill(oldfd: RawFd, newfd: RawFd, flags: OFlag) -> Result<RawFd> {
	if oldfd == newfd {
	return Err(Error::Sys(Errno::EINVAL));
	}
	let fd = try!(dup2(oldfd, newfd));
	if flags.contains(OFlag::O_CLOEXEC) {
	if let Err(e) = fcntl(fd, F_SETFD(FdFlag::FD_CLOEXEC)) {
	let _ = close(fd);
	return Err(e);
	}
	}
	Ok(fd)
	}

This means that using it in multithreaded applications results in races even on platforms where a proper dup3 is supported.

I think it would be best if it used dup3 if it is available and only falls back to dup2 if dup3 is not available.

[asomers]

Indeed. It looks like the actual dup3 syscall was removed in 9e93533 . But I would go further than you suggest. Our usual convention in Nix is to not even attempt to emulate nonexistent syscalls. Of course, we don't want to remove things that people may be using. See pipe2 in src/unistd.rs. Would you be willing to try fixing this?