M10 Phase 6: Complete Userspace Networking

Phase 4 wired userspace tools to the kernel's smoltcp network stack — ifconfig worked, DNS config was in place, OpenRC's networking service came up clean. But wget and curl still couldn't connect. The problem: both tools use nonblocking connect with poll/select for timeout handling, and our TCP connect always blocked.

Nonblocking connect

The existing TcpSocket::connect() ignored the options parameter entirely. It called sleep_signalable_until() waiting for may_send() to become true, regardless of whether O_NONBLOCK was set.

The fix follows the POSIX/Linux model:

1. Initiate the TCP SYN via smoltcp
2. If nonblocking, return EINPROGRESS immediately
3. The caller polls for POLLOUT (connection established) or POLLERR (connection failed)
4. getsockopt(SO_ERROR) reports the result

#![allow(unused)]
fn main() {
fn connect(&self, sockaddr: SockAddr, options: &OpenOptions) -> Result<()> {
    // ... SYN initiation, unchanged ...
    process_packets();

    if options.nonblock {
        return Err(Errno::EINPROGRESS.into());
    }

    // Blocking path now checks state properly
    SOCKET_WAIT_QUEUE.sleep_signalable_until(|| {
        let socket: &tcp::Socket = sockets.get(self.handle);
        match socket.state() {
            tcp::State::Established => Ok(Some(())),
            tcp::State::Closed => Err(Errno::ECONNREFUSED.into()),
            _ => Ok(None),
        }
    })
}
}

The blocking path also improved: previously it checked may_send() which doesn't distinguish "still connecting" from "connection failed". Now it inspects smoltcp's TCP state machine directly — Established means success, Closed means the remote sent RST (ECONNREFUSED).

Two guard checks at the top handle re-entrant connect calls: EISCONN if already established, EALREADY if a SYN is already in flight. Both are required by POSIX and expected by wget/curl.

SO_ERROR with real state

The old getsockopt(SO_ERROR) always returned 0. After a nonblocking connect, the caller needs to know whether the connection succeeded or failed. The new implementation polls the socket — if POLLERR is set (which TcpSocket::poll() now reports for State::Closed), it returns ECONNREFUSED (111).

This completes the nonblocking connect lifecycle:

socket() → fcntl(O_NONBLOCK) → connect() = EINPROGRESS
→ poll(POLLOUT) → getsockopt(SO_ERROR) = 0  (success)

ICMP ping socket

BusyBox ping uses Linux's "ping socket" feature: socket(AF_INET, SOCK_DGRAM, IPPROTO_ICMP). This avoids raw sockets (which require root) by letting the kernel handle ICMP echo request/reply framing.

The new IcmpSocket wraps smoltcp's icmp::Socket:

• Auto-bind: Generates a random ICMP identifier on first send (BusyBox doesn't call bind() on ping sockets)
• sendto: Writes raw ICMP bytes to smoltcp's transmit buffer, addressed to the destination IP
• recvfrom: Returns ICMP reply bytes with the source address as a sockaddr_in

Required adding socket-icmp to smoltcp's feature flags in Cargo.toml.

Everything else

New errnos: EINPROGRESS (115) and EALREADY (114) added to the Errno enum.

SO_RCVTIMEO/SO_SNDTIMEO: wget and curl set receive/send timeouts via setsockopt. Accepted silently — signal interruption via EINTR already provides the timeout escape hatch.

getsockopt stubs: SO_RCVBUF returns 87380, SO_SNDBUF returns 16384, SO_KEEPALIVE returns 0. Reasonable defaults that satisfy probing by networking tools.

/proc/net/ stubs: Added /proc/net/tcp, /proc/net/udp, /proc/net/tcp6, /proc/net/udp6 — each returns just the header line. Some libraries and tools check these exist.

Files changed