[Regression] Local DNS now defaults to tcp_only mode, causing UDP-only clients to fail

[spyophobia]

Steps to reproduce

cargo test --features dns-over-https,dns-over-tls,local-dns,local-http-rustls,local-redir,local-tun dns_relay -- --nocapture

Observe that the test hangs indefinitely.

Environments tested

1. My local machine: x86_64, ArchLinux, Rust 1.72.0 (x86_64-unknown-linux-gnu)
2. My COPR build: x86_64 & aarch64, RHEL 7-9 & Fedora 37-Rawhide (containerized), Rust 1.72.0 (x86_64-unknown-linux-gnu / aarch64-unknown-linux-gnu)

Things tried

1. Disable dns_relay test: cargo test --features dns-over-https,dns-over-tls,local-dns,local-http-rustls,local-redir,local-tun -- --skip dns_relay; all other tests complete successfully
2. Disable firewall: sudo systemctl stop firewalld; no effect

Debugging info

• I inserted some printlns and it seems like it's stuck here:

shadowsocks-rust/tests/dns.rs

Line 88 in b8ea553

let n = c.recv(&mut buf).await.unwrap();

• A git bisect pinpointed the issue to 8d3461b.

[spyophobia]

Two issues here:

1. Test is failing.
2. Can we institute some kind of timeout for all tests? This way the tests fail instead of hanging forever. Unfortunately cargo doesn't yet support timeouts (rfc-2798), so maybe use something like ntest?

[spyophobia]

I just noticed today that local DNS is just not working at all in 1.16.0. Connections are all getting refused. So this is a serious regression. I'll take a look at the problematic commit and see what I can do.

[spyophobia]

Ah okay. After some investigation it seems like the issue is less severe than I thought.

So in 1.15.4, the local DNS client will always create a TCP and a UDP server regardless of the mode configured:

shadowsocks-rust/crates/shadowsocks-service/src/local/dns/server.rs

Lines 76 to 88 in 16f280c

	pub async fn run(self, bind_addr: &ServerAddr, balancer: PingBalancer) -> io::Result<()> {
	let client = Arc::new(DnsClient::new(self.context.clone(), balancer, self.mode));
	let tcp_fut = self.run_tcp_server(bind_addr, client.clone());
	let udp_fut = self.run_udp_server(bind_addr, client);
	tokio::pin!(tcp_fut, udp_fut);
	match future::select(tcp_fut, udp_fut).await {
	Either::Left((res, ..)) => res,
	Either::Right((res, ..)) => res,
	}
	}

But in 1.16.0, the mode is now respected:

shadowsocks-rust/crates/shadowsocks-service/src/local/dns/server.rs

Lines 100 to 117 in 5f8fd17

	let mut tcp_server = None;
	if self.mode.enable_tcp() {
	let server = DnsTcpServer::new(
	self.context.clone(),
	&self.bind_addr,
	local_addr.clone(),
	remote_addr.clone(),
	client.clone(),
	)
	.await?;
	tcp_server = Some(server);
	}
	let mut udp_server = None;
	if self.mode.enable_udp() {
	let server = DnsUdpServer::new(self.context, &self.bind_addr, local_addr, remote_addr, client).await?;
	udp_server = Some(server);
	}

This is a good change in theory; unfortunately it had the side-effect of changing the default behavior from "TCP and UDP" to "TCP only", and AFAIK, was not documented in the changelog. What's worse is that lots of DNS clients only support UDP (in particular, Windows; this is how I initially noticed the problem; my Windows VM was having DNS problems while my Linux VMs were fine), which made it look like the feature has stopped working entirely.