bxcore, loopd, inferproxy: a lock file keeps each to one instance

Co-Authored-By: ornith-1.5-35b-a3b <agent@wntrmute.dev>

crates/bxcore/src/instance.rs

@@ -0,0 +1,97 @@
+//! One instance per daemon (M5b spec, section 9a). Every daemon removes a socket it finds at its
+//! path before it binds, so a second copy started by mistake would take the sockets of the one
+//! that runs. Before it touches them, a daemon takes an exclusive lock on its lock file with
+//! `lock`, and keeps the `InstanceLock` until the process ends.
+
+use std::fs::{DirBuilder, File, OpenOptions, TryLockError};
+use std::os::unix::fs::{DirBuilderExt, OpenOptionsExt};
+use std::path::Path;
+
+/// The runbook entry for both ways `lock` can fail.
+pub const RUNBOOK: &str = "see docs/runbook.md#already-running";
+
+/// A held lock: the open lock file. The system releases the lock when the file is closed, which is
+/// when this is dropped or the process ends, however it ends. The file itself stays: removing it
+/// would let a second copy lock a new file while the first still runs.
+#[derive(Debug)]
+#[must_use = "the lock is released when this value is dropped"]
+pub struct InstanceLock {
+    _file: File,
+}
+
+/// Why `lock` did not take the lock.
+#[derive(Debug)]
+pub enum LockError {
+    /// Another open file holds the lock: another instance is running.
+    Held,
+    /// The lock file, or its directory, could not be created or opened, or the lock failed.
+    Io(std::io::Error),
+}
+
+impl std::fmt::Display for LockError {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        match self {
+            LockError::Held => write!(f, "another instance holds the lock"),
+            LockError::Io(e) => write!(f, "{e}"),
+        }
+    }
+}
+
+impl std::error::Error for LockError {
+    fn source(&self) -> Option<&(dyn std::error::Error + 'static)> {
+        match self {
+            LockError::Held => None,
+            LockError::Io(e) => Some(e),
+        }
+    }
+}
+
+impl LockError {
+    /// The line a daemon prints when it does not start: `role` is its name, `path` the lock file.
+    pub fn line(&self, role: &str, path: &Path) -> String {
+        // `Held`: "{role}: another {role} is running (it holds {path}); {RUNBOOK}".
+        // `Io(e)`: "{role}: cannot take the lock {path}: {e}; {RUNBOOK}".
+        // Each with `format!`, the path shown with `path.display()`.
+        match self {
+            LockError::Held => format!(
+                "{role}: another {role} is running (it holds {}); {RUNBOOK}",
+                path.display()
+            ),
+            LockError::Io(e) => format!(
+                "{role}: cannot take the lock {path}: {e}; {RUNBOOK}",
+                path = path.display()
+            ),
+        }
+    }
+}
+
+/// Take the exclusive lock on `path`, creating the file with mode 0600, and its directory with
+/// mode 0700 when it is missing, as the daemons make their socket directories. The lock is tried
+/// once, never waited for.
+pub fn lock(path: &Path) -> Result<InstanceLock, LockError> {
+    // 1. If `path.parent()` is `Some(dir)`: `DirBuilder::new().recursive(true).mode(0o700)
+    //    .create(dir)`. 2. `OpenOptions::new().write(true).create(true).truncate(false)
+    //    .mode(0o600).open(path)`. An error of either is `LockError::Io`, through `map_err`.
+    // 3. `file.try_lock()`: `Ok(())` is `Ok(InstanceLock { _file: file })`,
+    //    `Err(TryLockError::WouldBlock)` is `Err(LockError::Held)`, and
+    //    `Err(TryLockError::Error(e))` is `Err(LockError::Io(e))`.
+    if let Some(dir) = path.parent() {
+        DirBuilder::new()
+            .recursive(true)
+            .mode(0o700)
+            .create(dir)
+            .map_err(LockError::Io)?;
+    }
+    let file = OpenOptions::new()
+        .write(true)
+        .create(true)
+        .truncate(false)
+        .mode(0o600)
+        .open(path)
+        .map_err(LockError::Io)?;
+    match file.try_lock() {
+        Ok(()) => Ok(InstanceLock { _file: file }),
+        Err(TryLockError::WouldBlock) => Err(LockError::Held),
+        Err(TryLockError::Error(e)) => Err(LockError::Io(e)),
+    }
+}

crates/bxcore/src/lib.rs

@@ -1,5 +1,6 @@
 //! Shared data types and the frame codec for Boxmaker. No policy and no I/O beyond frames, except
-//! the secret store (`secrets`), behind the feature of that name.
+//! the secret store (`secrets`), behind the feature of that name, and the lock that keeps a daemon
+//! to one instance (`instance`).
 
 pub mod audit;
 pub mod chain;
@@ -11,6 +12,7 @@ pub mod grant;
 pub mod hash;
 pub mod hosts;
 pub mod ids;
+pub mod instance;
 pub mod log;
 pub mod memory;
 pub mod sanitize;

crates/bxcore/tests/instance.rs

@@ -0,0 +1,137 @@
+//! `bxcore::instance`: the lock that keeps a daemon to one instance (M5b spec, section 9a). Two
+//! opens of the same file in one process are two holders, as two processes are. Do not edit.
+
+#[path = "support/tmp.rs"]
+mod tmp;
+
+use std::os::unix::fs::PermissionsExt;
+use std::path::{Path, PathBuf};
+
+use bxcore::instance::{LockError, RUNBOOK, lock};
+use tmp::TempDir;
+
+/// Gives a directory back its owner's permissions when dropped, so the `TempDir` declared before
+/// it can be removed even when an assertion fails (tip T44).
+struct Restore(PathBuf);
+
+impl Drop for Restore {
+    fn drop(&mut self) {
+        let _ = std::fs::set_permissions(&self.0, std::fs::Permissions::from_mode(0o700));
+    }
+}
+
+fn mode(path: &Path) -> u32 {
+    std::fs::metadata(path).unwrap().permissions().mode() & 0o777
+}
+
+#[test]
+fn a_second_lock_is_held_while_the_first_lives() {
+    let dir = TempDir::new("lock-held");
+    let path = dir.path().join("run").join("loopd.lock");
+    let first = lock(&path).expect("the first lock");
+    for _ in 0..3 {
+        match lock(&path) {
+            Err(LockError::Held) => {}
+            other => panic!("a second lock while the first lives: {other:?}"),
+        }
+    }
+    drop(first);
+}
+
+#[test]
+fn the_lock_is_free_again_once_the_first_is_dropped() {
+    let dir = TempDir::new("lock-free");
+    let path = dir.path().join("run").join("loopd.lock");
+    let first = lock(&path).expect("the first lock");
+    drop(first);
+    let second = lock(&path).expect("the lock after the first was dropped");
+    assert!(
+        matches!(lock(&path), Err(LockError::Held)),
+        "the second holds it now"
+    );
+    drop(second);
+    assert!(path.exists(), "the lock file is left in place");
+}
+
+#[test]
+fn a_lock_file_left_behind_is_taken() {
+    // A killed daemon leaves its lock file, unlocked; the next start takes it and keeps its bytes.
+    let dir = TempDir::new("lock-left");
+    let path = dir.write("run/memoryd.lock", "left\n");
+    let held = lock(&path).expect("a lock file nobody holds");
+    assert_eq!(std::fs::read_to_string(&path).unwrap(), "left\n");
+    drop(held);
+}
+
+#[test]
+fn the_lock_file_is_created_with_mode_0600_and_its_directory_0700() {
+    let dir = TempDir::new("lock-mode");
+    let run = dir.path().join("home").join("run");
+    let path = run.join("brokerd.lock");
+    let held = lock(&path).expect("a lock in a directory that does not exist yet");
+    assert_eq!(mode(&path), 0o600, "the lock file");
+    assert_eq!(mode(&run), 0o700, "the directory it made");
+    assert_eq!(mode(&dir.path().join("home")), 0o700, "and the one above");
+    assert_eq!(
+        std::fs::metadata(&path).unwrap().len(),
+        0,
+        "nothing is written"
+    );
+    drop(held);
+}
+
+#[test]
+fn a_directory_that_cannot_be_written_is_an_io_error() {
+    let dir = TempDir::new("lock-ro");
+    let run = dir.path().join("run");
+    std::fs::create_dir(&run).unwrap();
+    let _restore = Restore(run.clone());
+    std::fs::set_permissions(&run, std::fs::Permissions::from_mode(0o500)).unwrap();
+    match lock(&run.join("gatewayd.lock")) {
+        Err(LockError::Io(e)) => {
+            assert_eq!(e.kind(), std::io::ErrorKind::PermissionDenied, "{e}");
+        }
+        other => panic!("a lock in a read-only directory: {other:?}"),
+    }
+    assert!(!run.join("gatewayd.lock").exists());
+}
+
+#[test]
+fn a_path_through_a_file_is_an_io_error() {
+    let dir = TempDir::new("lock-file");
+    let file = dir.write("run", "not a directory\n");
+    match lock(&file.join("loopd.lock")) {
+        Err(LockError::Io(_)) => {}
+        other => panic!("a lock under a file: {other:?}"),
+    }
+}
+
+#[test]
+fn the_lines_a_daemon_prints() {
+    let path = Path::new("/home/bx/run/loopd.lock");
+    assert_eq!(
+        LockError::Held.line("loopd", path),
+        "loopd: another loopd is running (it holds /home/bx/run/loopd.lock); \
+         see docs/runbook.md#already-running"
+    );
+    let io = LockError::Io(std::io::Error::other("the disk is on fire"));
+    assert_eq!(
+        io.line("inferproxy", Path::new("/run/bx/infer.sock.lock")),
+        "inferproxy: cannot take the lock /run/bx/infer.sock.lock: the disk is on fire; \
+         see docs/runbook.md#already-running"
+    );
+    assert_eq!(RUNBOOK, "see docs/runbook.md#already-running");
+}
+
+#[test]
+fn the_error_displays_and_has_a_source() {
+    use std::error::Error;
+    assert_eq!(
+        LockError::Held.to_string(),
+        "another instance holds the lock"
+    );
+    assert!(LockError::Held.source().is_none());
+    let io = LockError::Io(std::io::Error::other("no room"));
+    assert_eq!(io.to_string(), "no room");
+    assert!(io.source().is_some());
+}

crates/inferproxy/src/main.rs

@@ -4,6 +4,7 @@ use std::os::unix::fs::PermissionsExt;
 use std::path::PathBuf;
 use std::process;
 
+use bxcore::instance;
 use inferproxy::{Limits, serve};
 
 fn main() {
@@ -39,6 +40,19 @@ fn main() {
     };
 
     let path: PathBuf = listen.into();
+    // One inferproxy per socket (M5b spec, section 9a): the lock beside the socket comes before
+    // the stale socket is removed, so a second copy never takes the socket of the one that runs.
+    // Held until the process ends; `process::exit` leaves it to the system to release.
+    let mut lock_path = path.clone().into_os_string();
+    lock_path.push(".lock");
+    let lock_path = PathBuf::from(lock_path);
+    let _instance = match instance::lock(&lock_path) {
+        Ok(held) => held,
+        Err(e) => {
+            eprintln!("{}", e.line("inferproxy", &lock_path));
+            process::exit(1);
+        }
+    };
     match fs::remove_file(&path) {
         Ok(()) => {}
         Err(e) if e.kind() == std::io::ErrorKind::NotFound => {}

crates/inferproxy/tests/instance.rs

@@ -0,0 +1,228 @@
+//! One `inferproxy` per socket (M5b spec, section 9a): the lock is `<socket>.lock`, beside the
+//! `--listen` socket. A second `inferproxy` on the same socket is refused, and the first keeps
+//! forwarding; once the first is killed, a new one removes the stale socket and forwards. Every
+//! wait has a deadline, and every process is killed and every path removed by a guard (tip T44).
+//! Do not edit.
+
+use std::io::{Read, Write};
+use std::net::TcpListener;
+use std::os::unix::fs::PermissionsExt;
+use std::os::unix::net::UnixStream;
+use std::path::{Path, PathBuf};
+use std::process::{Child, Command, Stdio};
+use std::sync::atomic::{AtomicU32, Ordering};
+use std::thread;
+use std::time::{Duration, Instant};
+
+const POINTER: &str = "see docs/runbook.md#already-running";
+
+static NEXT: AtomicU32 = AtomicU32::new(0);
+
+/// A temporary directory for the socket, removed when dropped, first with its owner's permissions
+/// put back.
+struct Dir(PathBuf);
+
+impl Dir {
+    fn new() -> Dir {
+        let n = NEXT.fetch_add(1, Ordering::SeqCst);
+        let dir = std::env::temp_dir().join(format!("ip-lock-{}-{n}", std::process::id()));
+        let _ = std::fs::remove_dir_all(&dir);
+        std::fs::create_dir_all(&dir).unwrap();
+        Dir(dir)
+    }
+
+    fn socket(&self) -> PathBuf {
+        self.0.join("infer.sock")
+    }
+}
+
+impl Drop for Dir {
+    fn drop(&mut self) {
+        let _ = std::fs::set_permissions(&self.0, std::fs::Permissions::from_mode(0o700));
+        let _ = std::fs::remove_dir_all(&self.0);
+    }
+}
+
+/// An upstream that reads four bytes from each connection, answers `pong` and closes.
+fn upstream() -> String {
+    let listener = TcpListener::bind("127.0.0.1:0").unwrap();
+    let addr = listener.local_addr().unwrap().to_string();
+    thread::spawn(move || {
+        for stream in listener.incoming() {
+            let Ok(mut stream) = stream else { break };
+            thread::spawn(move || {
+                let mut ping = [0u8; 4];
+                if stream.read_exact(&mut ping).is_ok() {
+                    let _ = stream.write_all(b"pong");
+                }
+            });
+        }
+    });
+    addr
+}
+
+/// A running `inferproxy`, killed and waited for when dropped. Declared after the `Dir`, so it
+/// drops first.
+struct Proxy(Child);
+
+impl Drop for Proxy {
+    fn drop(&mut self) {
+        let _ = self.0.kill();
+        let _ = self.0.wait();
+    }
+}
+
+impl Proxy {
+    fn running(&mut self) -> bool {
+        matches!(self.0.try_wait(), Ok(None))
+    }
+}
+
+fn command(socket: &Path, upstream: &str) -> Command {
+    let mut command = Command::new(env!("CARGO_BIN_EXE_inferproxy"));
+    command
+        .arg("--listen")
+        .arg(socket)
+        .args(["--upstream", upstream])
+        .stdin(Stdio::null())
+        .stdout(Stdio::null());
+    command
+}
+
+fn start(socket: &Path, upstream: &str) -> Proxy {
+    Proxy(
+        command(socket, upstream)
+            .stderr(Stdio::null())
+            .spawn()
+            .unwrap(),
+    )
+}
+
+/// Runs `inferproxy` to its end, at most 10 s (then it is killed), and gives its exit code
+/// (`None` if it had to be killed) and its stderr.
+fn run_to_end(socket: &Path, upstream: &str) -> (Option<i32>, String) {
+    let mut child = Proxy(
+        command(socket, upstream)
+            .stderr(Stdio::piped())
+            .spawn()
+            .unwrap(),
+    );
+    let deadline = Instant::now() + Duration::from_secs(10);
+    let mut code = None;
+    while Instant::now() < deadline {
+        if let Ok(Some(status)) = child.0.try_wait() {
+            code = status.code();
+            break;
+        }
+        thread::sleep(Duration::from_millis(20));
+    }
+    let _ = child.0.kill();
+    let _ = child.0.wait();
+    let mut stderr = String::new();
+    if let Some(mut pipe) = child.0.stderr.take() {
+        let _ = pipe.read_to_string(&mut stderr);
+    }
+    (code, stderr)
+}
+
+/// True when a `ping` sent to `socket` comes back from the upstream as `pong` within 5 s.
+fn forwards(socket: &Path) -> bool {
+    let Ok(mut stream) = UnixStream::connect(socket) else {
+        return false;
+    };
+    let limit = Some(Duration::from_secs(5));
+    if stream.set_read_timeout(limit).is_err() || stream.set_write_timeout(limit).is_err() {
+        return false;
+    }
+    if stream.write_all(b"ping").is_err() {
+        return false;
+    }
+    let mut answer = Vec::new();
+    let _ = stream.read_to_end(&mut answer);
+    answer == b"pong"
+}
+
+/// Waits up to 10 s for an `inferproxy` to forward on `socket`.
+fn wait_until_it_forwards(socket: &Path) -> bool {
+    let deadline = Instant::now() + Duration::from_secs(10);
+    while Instant::now() < deadline {
+        if forwards(socket) {
+            return true;
+        }
+        thread::sleep(Duration::from_millis(50));
+    }
+    false
+}
+
+fn inode(path: &Path) -> u64 {
+    use std::os::unix::fs::MetadataExt;
+    std::fs::symlink_metadata(path)
+        .map(|m| m.ino())
+        .unwrap_or(0)
+}
+
+#[test]
+fn a_second_inferproxy_on_the_same_socket_is_refused_and_the_first_keeps_forwarding() {
+    let dir = Dir::new();
+    let socket = dir.socket();
+    let upstream = upstream();
+    let mut first = start(&socket, &upstream);
+    assert!(wait_until_it_forwards(&socket), "the first never forwarded");
+    let before = inode(&socket);
+
+    let (code, stderr) = run_to_end(&socket, &upstream);
+    let line = format!(
+        "inferproxy: another inferproxy is running (it holds {}.lock); {POINTER}",
+        socket.display()
+    );
+    assert_eq!(code, Some(1), "the second inferproxy: {stderr}");
+    assert!(
+        stderr.lines().any(|l| l == line),
+        "wanted the line\n{line}\nstderr: {stderr}"
+    );
+
+    assert!(first.running(), "the first inferproxy is still running");
+    assert_eq!(inode(&socket), before, "the first's socket is still there");
+    assert!(forwards(&socket), "the first still forwards");
+}
+
+#[test]
+fn after_the_first_is_killed_a_new_inferproxy_removes_the_stale_socket_and_forwards() {
+    let dir = Dir::new();
+    let socket = dir.socket();
+    let upstream = upstream();
+    let mut first = start(&socket, &upstream);
+    assert!(wait_until_it_forwards(&socket), "the first never forwarded");
+    let _ = first.0.kill();
+    let _ = first.0.wait();
+    assert!(socket.exists(), "a killed inferproxy leaves its socket");
+    assert!(!forwards(&socket), "nobody serves the stale socket");
+
+    let _next = start(&socket, &upstream);
+    assert!(
+        wait_until_it_forwards(&socket),
+        "the new one never forwarded"
+    );
+    let lock = dir.0.join("infer.sock.lock");
+    let mode = std::fs::metadata(&lock).map(|m| m.permissions().mode() & 0o777);
+    assert_eq!(mode.ok(), Some(0o600), "the lock file, beside the socket");
+}
+
+#[test]
+fn a_socket_directory_that_cannot_be_written_stops_inferproxy_naming_the_error() {
+    let dir = Dir::new();
+    let socket = dir.socket();
+    std::fs::set_permissions(&dir.0, std::fs::Permissions::from_mode(0o500)).unwrap();
+
+    let (code, stderr) = run_to_end(&socket, &upstream());
+    let line = format!(
+        "inferproxy: cannot take the lock {}.lock: Permission denied (os error 13); {POINTER}",
+        socket.display()
+    );
+    assert_eq!(code, Some(1), "{stderr}");
+    assert!(
+        stderr.lines().any(|l| l == line),
+        "wanted the line\n{line}\nstderr: {stderr}"
+    );
+    assert!(!socket.exists(), "no socket");
+}

crates/loopd/src/main.rs

@@ -8,6 +8,7 @@ use std::process::ExitCode;
 use std::sync::Arc;
 use std::time::Duration;
 
+use bxcore::instance;
 use loopd::broker_port::{BrokerPort, NoBroker, not_configured_line};
 use loopd::channel::{self, Context};
 use loopd::config::Config;
@@ -73,6 +74,18 @@ fn run_serve(path: &Path) -> ExitCode {
         }
     };
 
+    // One loopd per home (M5b spec, section 9a). The lock comes before the socket is touched and
+    // before the self-test, so a second copy neither takes the socket of the one that runs nor
+    // sends its slot a prompt. Held until the process ends.
+    let lock_path = cfg.paths.home.join("run/loopd.lock");
+    let _instance = match instance::lock(&lock_path) {
+        Ok(held) => held,
+        Err(e) => {
+            eprintln!("{}", e.line("loopd", &lock_path));
+            return ExitCode::from(1);
+        }
+    };
+
     let socket = cfg.channel_socket();
     if socket.exists()
         && let Err(e) = std::fs::remove_file(&socket)

crates/loopd/tests/instance.rs

@@ -0,0 +1,311 @@
+//! One `loopd` per home (M5b spec, section 9a): a second `loopd serve` against the same home is
+//! refused before it touches the socket or the server, and the first keeps serving; once the first
+//! is killed, a new one removes the stale socket and serves. Every wait has a deadline, and every
+//! process is killed and every path removed by a guard (tip T44). Do not edit.
+
+mod support;
+
+use std::os::unix::fs::PermissionsExt;
+use std::os::unix::net::UnixStream;
+use std::path::{Path, PathBuf};
+use std::process::{Child, Command, Stdio};
+use std::time::{Duration, Instant};
+
+use bxcore::{
+    Envelope, ErrorCode, Message, PROTOCOL_VERSION, SessionId, Turn, WireError, read_frame,
+    write_frame,
+};
+use support::{FakeServer, Home, Reply, tokenizer};
+
+const POINTER: &str = "see docs/runbook.md#already-running";
+
+fn config_file(home: &Home, server: &FakeServer) -> PathBuf {
+    let text = format!(
+        r#"
+[infer]
+socket = "{}"
+model = "test-model"
+[slots]
+main = 0
+background = 1
+[expect]
+template_sha256 = "f55f52930aa8bf44ab5cb85f99370fcc3c56e9a85640b812086d5330bce5d86b"
+n_ctx = 131072
+slots = 2
+[limits]
+poll_ms = 40
+liveness_ms = 500
+retry_backoff_ms = [10]
+[paths]
+home = "{}"
+"#,
+        server.socket.display(),
+        home.dir.display()
+    );
+    let path = home.dir.join("config.toml");
+    std::fs::write(&path, text).unwrap();
+    path
+}
+
+/// A server that passes the self-test twice: once for the first `loopd`, once for the one started
+/// after it is killed.
+fn healthy_routes(server: &FakeServer) {
+    server.route("/props", vec![Reply::fixture("props")]);
+    server.respond("/tokenize", tokenizer);
+    server.route(
+        "/v1/chat/completions",
+        vec![
+            Reply::fixture("tool_call"),
+            Reply::fixture("turn1"),
+            Reply::fixture("turn2"),
+            Reply::fixture("tool_call"),
+            Reply::fixture("turn1"),
+            Reply::fixture("turn2"),
+        ],
+    );
+}
+
+fn socket(home: &Home) -> PathBuf {
+    home.dir.join("run").join("loop").join("loop.sock")
+}
+
+/// A running `loopd serve`, killed and waited for when dropped. Declared after the `Home` and the
+/// `FakeServer`, so it drops first.
+struct Daemon(Child);
+
+impl Drop for Daemon {
+    fn drop(&mut self) {
+        let _ = self.0.kill();
+        let _ = self.0.wait();
+    }
+}
+
+impl Daemon {
+    fn running(&mut self) -> bool {
+        matches!(self.0.try_wait(), Ok(None))
+    }
+}
+
+/// `loopd serve` with its stderr in `<home>/<name>.err`, for the messages of a failing test.
+fn start(home: &Home, config: &Path, name: &str) -> Daemon {
+    let err = std::fs::File::create(home.dir.join(format!("{name}.err"))).unwrap();
+    let child = Command::new(env!("CARGO_BIN_EXE_loopd"))
+        .args(["serve", "--config"])
+        .arg(config)
+        .stdin(Stdio::null())
+        .stdout(Stdio::null())
+        .stderr(Stdio::from(err))
+        .spawn()
+        .unwrap();
+    Daemon(child)
+}
+
+fn stderr_of(home: &Home, name: &str) -> String {
+    std::fs::read_to_string(home.dir.join(format!("{name}.err"))).unwrap_or_default()
+}
+
+/// Runs `loopd serve` to its end, at most 20 s (then it is killed), and gives its exit code
+/// (`None` if it had to be killed) and its stderr.
+fn run_to_end(config: &Path) -> (Option<i32>, String) {
+    let mut child = Daemon(
+        Command::new(env!("CARGO_BIN_EXE_loopd"))
+            .args(["serve", "--config"])
+            .arg(config)
+            .stdin(Stdio::null())
+            .stdout(Stdio::null())
+            .stderr(Stdio::piped())
+            .spawn()
+            .unwrap(),
+    );
+    let deadline = Instant::now() + Duration::from_secs(20);
+    let mut code = None;
+    while Instant::now() < deadline {
+        if let Ok(Some(status)) = child.0.try_wait() {
+            code = status.code();
+            break;
+        }
+        std::thread::sleep(Duration::from_millis(20));
+    }
+    let _ = child.0.kill();
+    let _ = child.0.wait();
+    let mut stderr = String::new();
+    if let Some(mut pipe) = child.0.stderr.take() {
+        let _ = std::io::Read::read_to_string(&mut pipe, &mut stderr);
+    }
+    (code, stderr)
+}
+
+/// True when `loopd` answers on `socket`: a turn that resumes a session that does not exist gets
+/// the error frame for it. Any other answer, or none within 5 s, is false.
+fn answers(socket: &Path) -> bool {
+    let Ok(mut stream) = UnixStream::connect(socket) else {
+        return false;
+    };
+    let limit = Some(Duration::from_secs(5));
+    if stream.set_read_timeout(limit).is_err() || stream.set_write_timeout(limit).is_err() {
+        return false;
+    }
+    let turn = Envelope {
+        v: PROTOCOL_VERSION,
+        id: 7,
+        r#final: true,
+        msg: Message::Turn(Turn {
+            session: SessionId::new("probe").unwrap(),
+            content: "are you there".to_string(),
+            resume: true,
+        }),
+    };
+    if write_frame(&mut stream, &turn).is_err() {
+        return false;
+    }
+    matches!(
+        read_frame(&mut stream),
+        Ok(Envelope {
+            msg: Message::Error(WireError {
+                code: ErrorCode::NoSuchSession,
+                ..
+            }),
+            ..
+        })
+    )
+}
+
+/// Waits up to 20 s for `loopd` to answer on `socket`.
+fn wait_until_it_answers(socket: &Path) -> bool {
+    let deadline = Instant::now() + Duration::from_secs(20);
+    while Instant::now() < deadline {
+        if answers(socket) {
+            return true;
+        }
+        std::thread::sleep(Duration::from_millis(50));
+    }
+    false
+}
+
+fn inode(path: &Path) -> u64 {
+    use std::os::unix::fs::MetadataExt;
+    std::fs::symlink_metadata(path)
+        .map(|m| m.ino())
+        .unwrap_or(0)
+}
+
+#[test]
+fn a_second_loopd_on_the_same_home_is_refused_and_the_first_keeps_serving() {
+    let home = Home::new();
+    let server = FakeServer::start();
+    healthy_routes(&server);
+    let config = config_file(&home, &server);
+    let socket = socket(&home);
+    let mut first = start(&home, &config, "first");
+    assert!(
+        wait_until_it_answers(&socket),
+        "the first loopd never answered; stderr: {}",
+        stderr_of(&home, "first")
+    );
+    let before = inode(&socket);
+    let completions = server.requests_to("/v1/chat/completions").len();
+    let tokenized = server.requests_to("/tokenize").len();
+
+    let (code, stderr) = run_to_end(&config);
+    let lock = home.dir.join("run").join("loopd.lock");
+    let line = format!(
+        "loopd: another loopd is running (it holds {}); {POINTER}",
+        lock.display()
+    );
+    assert_eq!(code, Some(1), "the second loopd: {stderr}");
+    assert!(
+        stderr.lines().any(|l| l == line),
+        "wanted the line\n{line}\nstderr: {stderr}"
+    );
+    assert!(
+        !stderr.contains("selftest"),
+        "the second ran no self-test: {stderr}"
+    );
+    assert_eq!(
+        (
+            server.requests_to("/v1/chat/completions").len(),
+            server.requests_to("/tokenize").len()
+        ),
+        (completions, tokenized),
+        "the second sent the server nothing"
+    );
+
+    assert!(first.running(), "the first loopd is still running");
+    assert_eq!(inode(&socket), before, "the first's socket is still there");
+    assert!(
+        answers(&socket),
+        "the first loopd still answers; its stderr: {}",
+        stderr_of(&home, "first")
+    );
+}
+
+#[test]
+fn after_the_first_is_killed_a_new_loopd_removes_the_stale_socket_and_serves() {
+    let home = Home::new();
+    let server = FakeServer::start();
+    healthy_routes(&server);
+    let config = config_file(&home, &server);
+    let socket = socket(&home);
+    let mut first = start(&home, &config, "first");
+    assert!(
+        wait_until_it_answers(&socket),
+        "the first loopd never answered; stderr: {}",
+        stderr_of(&home, "first")
+    );
+    let _ = first.0.kill();
+    let _ = first.0.wait();
+    assert!(socket.exists(), "a killed loopd leaves its socket");
+    assert!(!answers(&socket), "nobody serves the stale socket");
+
+    let _next = start(&home, &config, "next");
+    assert!(
+        wait_until_it_answers(&socket),
+        "the new loopd never answered; stderr: {}",
+        stderr_of(&home, "next")
+    );
+    let lock = home.dir.join("run").join("loopd.lock");
+    let mode = std::fs::metadata(&lock).map(|m| m.permissions().mode() & 0o777);
+    assert_eq!(mode.ok(), Some(0o600), "the lock file, mode 0600");
+    assert_eq!(
+        server.requests_to("/v1/chat/completions").len(),
+        6,
+        "each of the two ran the self-test"
+    );
+}
+
+/// Gives `<home>/run` back its owner's permissions when dropped, before the `Home` is removed.
+struct Restore(PathBuf);
+
+impl Drop for Restore {
+    fn drop(&mut self) {
+        let _ = std::fs::set_permissions(&self.0, std::fs::Permissions::from_mode(0o700));
+    }
+}
+
+#[test]
+fn a_run_directory_that_cannot_be_written_stops_loopd_naming_the_error() {
+    let home = Home::new();
+    let server = FakeServer::start();
+    healthy_routes(&server);
+    let config = config_file(&home, &server);
+    let run = home.dir.join("run");
+    std::fs::create_dir(&run).unwrap();
+    let _restore = Restore(run.clone());
+    std::fs::set_permissions(&run, std::fs::Permissions::from_mode(0o500)).unwrap();
+
+    let (code, stderr) = run_to_end(&config);
+    let line = format!(
+        "loopd: cannot take the lock {}: Permission denied (os error 13); {POINTER}",
+        run.join("loopd.lock").display()
+    );
+    assert_eq!(code, Some(1), "{stderr}");
+    assert!(
+        stderr.lines().any(|l| l == line),
+        "wanted the line\n{line}\nstderr: {stderr}"
+    );
+    assert!(
+        server.requests().is_empty(),
+        "loopd sent the server nothing"
+    );
+    assert!(!socket(&home).exists(), "no socket");
+}