Files
galaxy/app/src/remote_server/unix/proxy.rs
T

377 lines
15 KiB
Rust

//! Remote server proxy — runs over SSH stdio and bridges to the long-lived
//! daemon process via a Unix domain socket.
//!
//! Responsibilities:
//! 1. Acquire an exclusive `flock` on the PID file to serialise concurrent
//! proxy starts (e.g. two tabs SSH-ing to the same host at the same time).
//! 2. Check whether the daemon is already running (`kill -0`).
//! 3. If not: spawn the daemon subcommand in a new session and wait for its
//! socket to appear.
//! 4. Connect to `server.sock` and bridge stdin/stdout to the socket using
//! the existing 4-byte length-prefixed frame format.
use std::fs::Permissions;
use std::os::unix::fs::PermissionsExt;
use std::os::unix::io::AsRawFd;
use std::path::PathBuf;
use std::process::Stdio;
use std::time::Duration;
use super::super::setup;
/// Path to the daemon's Unix domain socket, versioned on release channels.
pub(super) fn socket_path(identity_key: &str) -> PathBuf {
let dir = setup::remote_server_daemon_dir(identity_key);
let expanded = shellexpand::tilde(&dir).into_owned();
PathBuf::from(expanded).join(setup::daemon_socket_name())
}
/// Path to the daemon's PID file (also used as the flock target),
/// versioned on release channels.
pub(super) fn pid_path(identity_key: &str) -> PathBuf {
let dir = setup::remote_server_daemon_dir(identity_key);
let expanded = shellexpand::tilde(&dir).into_owned();
PathBuf::from(expanded).join(setup::daemon_pid_name())
}
/// Daemon directory for the given identity key (expanded, no tilde).
fn daemon_dir(identity_key: &str) -> PathBuf {
let dir = setup::remote_server_daemon_dir(identity_key);
let expanded = shellexpand::tilde(&dir).into_owned();
PathBuf::from(expanded)
}
/// Scans the identity-key daemon directory and removes socket/PID files
/// from previous daemon versions.
///
/// Old daemons are **not** killed — they may still be serving active
/// connections from an older Warp client. Removing their socket file
/// prevents new proxies from accidentally connecting to them, and the
/// daemon's built-in grace timer (10 min with no connections) will shut
/// it down naturally after the last client disconnects.
///
/// Errors are logged but do not prevent the proxy from proceeding.
fn cleanup_old_versions(identity_key: &str) {
let dir = daemon_dir(identity_key);
let current_socket = setup::daemon_socket_name();
let current_pid = setup::daemon_pid_name();
let entries = match std::fs::read_dir(&dir) {
Ok(entries) => entries,
Err(_) => return,
};
for entry in entries.flatten() {
let name = entry.file_name();
let Some(name_str) = name.to_str() else {
continue;
};
// Remove old PID files: server*.pid that aren't the current version.
if name_str.ends_with(".pid") && name_str.starts_with("server") && name_str != current_pid {
log::info!("Proxy: removing old PID file {name_str}");
let _ = std::fs::remove_file(entry.path());
}
// Remove old socket files: server*.sock that aren't the current version.
if name_str.ends_with(".sock")
&& name_str.starts_with("server")
&& name_str != current_socket
{
log::info!("Proxy: removing old socket file {name_str}");
let _ = std::fs::remove_file(entry.path());
}
}
}
/// Ensures the daemon directory exists with owner-only permissions.
pub(super) fn ensure_private_daemon_dir(path: &std::path::Path) -> anyhow::Result<()> {
std::fs::create_dir_all(path)?;
std::fs::set_permissions(path, Permissions::from_mode(0o700))?;
Ok(())
}
/// Maximum usable `sun_path` length for Unix domain sockets.
///
/// macOS has the strictest limit (104 bytes including null terminator,
/// 103 usable). We use 103 on all platforms so a single binary works
/// everywhere without per-target branching.
const SUN_PATH_MAX: usize = 103;
/// Entry point for `remote-server-proxy`.
///
/// Ensures the daemon is running, then bridges stdin/stdout to the daemon's
/// Unix socket for the lifetime of this SSH session.
pub fn run(identity_key: &str) -> anyhow::Result<()> {
let socket_path = socket_path(identity_key);
let pid_path = pid_path(identity_key);
// Guard against socket paths that exceed the sun_path limit.
// Without this check, UnixListener::bind fails silently in the
// daemon and the proxy times out after 10s with no actionable
// error. With hashed identity + version names the path should
// always fit, so hitting this guard indicates a new path component
// was added without budgeting for sun_path. The error surfaces in
// client telemetry (RemoteServerInitialization) and daemon logs.
let path_len = socket_path.as_os_str().len();
if path_len > SUN_PATH_MAX {
anyhow::bail!(
"daemon socket path is {path_len} bytes, which exceeds the \
sun_path limit of {SUN_PATH_MAX} bytes: {}",
socket_path.display()
);
}
// Ensure the parent directory exists.
if let Some(parent) = socket_path.parent() {
ensure_private_daemon_dir(parent)?;
}
// Clean up socket/PID files from previous daemon versions so stale
// daemons left over after autoupdate don't linger.
cleanup_old_versions(identity_key);
// ---- Acquire exclusive flock on the PID file --------------------------------
//
// This serialises concurrent proxy starts. If two tabs SSH in at the
// same time and both see "no daemon running", only one will succeed in
// forking a daemon; the other will block here, then connect to the one
// the first proxy started.
//
// The lock is released automatically when the File is dropped.
let pid_file = std::fs::OpenOptions::new()
.create(true)
.truncate(false)
.read(true)
.write(true)
.open(&pid_path)?;
let pid_fd = pid_file.as_raw_fd();
flock_wait(pid_fd, libc::LOCK_EX)?;
// ---- Check whether daemon is already running --------------------------------
let daemon_running = check_daemon_running(&pid_path);
if daemon_running {
log::info!("Proxy: reusing existing daemon");
} else {
log::info!("Proxy: no daemon running, will start one");
}
if !daemon_running {
// Remove any stale socket from a previous crash.
if socket_path.exists() {
let _ = std::fs::remove_file(&socket_path);
}
// Spawn the daemon in a new Unix session so it is detached from
// the SSH session. When SSH exits the OS sends SIGHUP to every
// process in the session's foreground process group. `setsid()`
// creates a new session for the child, so the daemon is not in
// SSH's process group and will not receive that signal.
let exe = std::env::current_exe()?;
let mut cmd = command::blocking::Command::new(&exe);
cmd.arg("remote-server-daemon")
.arg("--identity-key")
.arg(identity_key)
.stdin(Stdio::null())
.stdout(Stdio::null())
.stderr(Stdio::null());
// SAFETY: setsid(2) is async-signal-safe and has no side effects
// other than creating a new session. pre_exec closures run between
// fork and exec in the child process.
unsafe {
cmd.pre_exec(|| {
libc::setsid();
Ok(())
});
}
cmd.spawn()
.map_err(|e| anyhow::anyhow!("failed to spawn daemon: {e}"))?;
// Wait for the daemon's socket to appear before releasing the flock.
// Holding the lock here prevents a concurrent proxy from acquiring it,
// reading a stale PID file, and racing to spawn a second daemon.
wait_for_socket(&socket_path)?;
flock_wait(pid_fd, libc::LOCK_UN)?;
drop(pid_file);
} else {
// Daemon already running — release the flock and connect.
flock_wait(pid_fd, libc::LOCK_UN)?;
drop(pid_file);
}
// ---- Bridge stdin/stdout to the daemon socket --------------------------------
bridge_stdio_to_socket(&socket_path)
}
/// Returns true if the PID stored in `pid_path` belongs to a live process.
fn check_daemon_running(pid_path: &std::path::Path) -> bool {
let Ok(contents) = std::fs::read_to_string(pid_path) else {
return false;
};
let Ok(pid) = contents.trim().parse::<libc::pid_t>() else {
return false;
};
// kill(pid, 0) succeeds (returns 0) if the process exists and we can
// signal it; it fails with ESRCH if the process does not exist.
// SAFETY: sending signal 0 is always safe — it performs a permission
// check only and does not deliver an actual signal.
unsafe { libc::kill(pid, 0) == 0 }
}
/// Poll until the daemon's socket file appears or the timeout elapses.
///
/// After we spawn the daemon there is a race: the daemon needs time to bind
/// and listen on the socket before the proxy can connect to it. We poll
/// until the socket file is present rather than connecting immediately,
/// which would fail with "no such file" if the daemon hasn't started yet.
fn wait_for_socket(socket_path: &std::path::Path) -> anyhow::Result<()> {
const TIMEOUT: Duration = Duration::from_secs(10);
const POLL_INTERVAL: Duration = Duration::from_millis(20);
let start = instant::Instant::now();
while !socket_path.exists() {
if start.elapsed() >= TIMEOUT {
anyhow::bail!(
"timed out waiting for daemon socket at {}",
socket_path.display()
);
}
std::thread::sleep(POLL_INTERVAL);
}
log::info!("Proxy: daemon socket ready after {:?}", start.elapsed());
Ok(())
}
/// Calls `flock(2)` with the given operation, retrying on `EINTR`.
///
/// Blocking `flock(LOCK_EX)` can be interrupted by a signal before acquiring
/// the lock; ignoring the return value would cause the proxy to proceed
/// without actually holding the lock.
fn flock_wait(fd: std::os::unix::io::RawFd, operation: libc::c_int) -> anyhow::Result<()> {
loop {
// SAFETY: flock(2) is safe to call with a valid fd and a valid operation.
let ret = unsafe { libc::flock(fd, operation) };
if ret == 0 {
return Ok(());
}
let err = std::io::Error::last_os_error();
if err.raw_os_error() == Some(libc::EINTR) {
continue; // Interrupted by signal — retry.
}
return Err(anyhow::anyhow!("flock failed: {err}"));
}
}
/// Connect to the daemon's Unix socket and copy bytes bidirectionally between
/// stdin/stdout and the socket.
///
/// The proxy is protocol-agnostic — it forwards raw bytes without parsing the
/// length-prefixed framing. The framing is handled at the endpoints (Warp
/// client and daemon).
///
/// **Important**: the stdout direction uses a manual read→write→flush loop
/// instead of `io::copy` because `std::io::stdout()` wraps the fd in a
/// `LineWriter` that only flushes up to the last `\n` byte in each write.
/// For a binary protocol the trailing bytes after the last `0x0a` get stuck
/// in the internal `BufWriter` and are never flushed, causing the client to
/// hang forever waiting for complete messages.
///
/// **Shutdown coordination**: each direction explicitly
/// [`shutdown(Both)`s][Shutdown] the Unix socket when its copy loop
/// returns, which unblocks the other thread's read/write on the same
/// underlying socket. Without this, when the client SIGKILLs the local
/// `ssh ... remote-server-proxy` slave (e.g. on `ExitShell`), sshd
/// closes our stdin but the daemon has no reason to close its end of
/// the Unix socket, so the stdout thread sits forever in a blocking
/// read. That keeps the proxy alive with stdout still open, which
/// keeps the SSH channel half-closed on the server side, which in
/// turn keeps the client's `ssh` ControlMaster from exiting until
/// sshd's session cleanup eventually fires. Shutting the Unix socket
/// here makes teardown deterministic and independent of whatever the
/// daemon is doing.
///
/// [Shutdown]: std::net::Shutdown
fn bridge_stdio_to_socket(socket_path: &std::path::Path) -> anyhow::Result<()> {
use std::io::{Read, Write};
use std::net::Shutdown;
log::info!(
"Proxy: connecting to daemon socket at {}",
socket_path.display()
);
let stream = std::os::unix::net::UnixStream::connect(socket_path)?;
log::info!("Proxy: connected, bridging stdio");
// Each thread holds two clones: one it actively reads/writes, and
// one used solely to `shutdown(Both)` on exit so the peer thread's
// blocking call returns. `UnixStream::try_clone` shares the
// underlying socket, so `shutdown` on any clone tears down both
// directions for every clone.
let stream_for_t1 = stream.try_clone()?;
let stream_shutdown_for_t1 = stream.try_clone()?;
let stream_for_t2 = stream.try_clone()?;
let stream_shutdown_for_t2 = stream.try_clone()?;
drop(stream);
let t1 = std::thread::Builder::new()
.name("proxy-stdin-fwd".into())
.spawn(move || {
let result = std::io::copy(&mut std::io::stdin(), &mut &stream_for_t1);
match &result {
Ok(total) => log::info!(
"Proxy: stdin->socket copy ended ({total} bytes); \
shutting down socket to unblock peer"
),
Err(e) => log::info!(
"Proxy: stdin->socket copy errored ({e}); \
shutting down socket to unblock peer"
),
}
let _ = stream_shutdown_for_t1.shutdown(Shutdown::Both);
result
})?;
// Socket → stdout: flush after every write so that complete protocol
// frames reach the SSH tunnel without waiting for the `LineWriter`
// buffer to fill.
let t2 = std::thread::Builder::new()
.name("proxy-stdout-fwd".into())
.spawn(move || -> std::io::Result<u64> {
let mut stdout = std::io::stdout().lock();
let mut buf = [0u8; 8192];
let mut total = 0u64;
let result = loop {
let n = match (&stream_for_t2).read(&mut buf) {
Ok(0) => break Ok(total),
Ok(n) => n,
Err(e) => break Err(e),
};
if let Err(e) = stdout.write_all(&buf[..n]) {
break Err(e);
}
if let Err(e) = stdout.flush() {
break Err(e);
}
total += n as u64;
};
match &result {
Ok(total) => log::info!(
"Proxy: socket->stdout copy ended ({total} bytes); \
shutting down socket to unblock peer"
),
Err(e) => log::info!(
"Proxy: socket->stdout copy errored ({e}); \
shutting down socket to unblock peer"
),
}
let _ = stream_shutdown_for_t2.shutdown(Shutdown::Both);
result
})?;
let _ = t1.join();
let _ = t2.join();
log::info!("Proxy: bridge closed, exiting");
Ok(())
}