blivet

A correct, full-featured Unix daemon library and CLI for Rust.

Daemonizing a process correctly is deceptively hard: the double-fork dance, session detachment, signal and fd hygiene, and -- the part most libraries skip -- telling the launcher whether the daemon actually came up. Unlike daemon(3) or thin wrappers, blivet reports post-fork startup failures back to the launcher over a notification pipe, with sysexits.h exit codes. The shell, systemd, or a supervisor sees a real success or failure -- not a detached process that may have already died during init.

use blivet::{daemonize, DaemonConfig};

let mut config = DaemonConfig::new();
config.pidfile("/var/run/myapp.pid");

let mut ctx = daemonize(&config)?; // safe: checks single-threaded, then double-forks
ctx.notify_parent()?;              // tell the launcher we're up; it exits 0
// daemon runs here

Why blivet

• Correct by default. Mandatory double-fork, setsid, signal reset (including real-time signals on Linux; SIGPIPE stays ignored so pipe writes keep returning errors instead of killing the daemon), signal-mask clear, fd close, and /dev/null redirect -- the things hand-rolled daemonizers forget.
• Parent notification. The launcher blocks until the daemon signals readiness or reports an error -- no "did it start?" polling. Forget to call notify_parent() and the launcher exits non-zero automatically.
• Split-phase privileges. daemonize() returns while still privileged, so you can bind port 80 or chroot before drop_privileges().
• Safe by default. The checked entry points verify single-threadedness, so you write no unsafe; the crate's own unsafe (libc FFI, fork, setenv) is isolated under #![deny(unsafe_code)] and documented. Opt-out variants are there when you manage the single-threaded contract yourself.
• Library and CLI. A builder API, or the standalone daemonize binary (cargo install blivet) that wraps any program.

How it works

daemonize() double-forks, calls setsid to detach from the controlling terminal, and returns a DaemonContext in the grandchild -- your daemon. The launcher does not return; it blocks on a pipe to the grandchild, waiting for that daemon to report in. In the daemon you then:

1. run fallible init -- bind sockets, open files, connect to dependencies;
2. chown_paths(), then drop_privileges(), to drop to an unprivileged user;
3. call notify_parent(), which writes one readiness byte down the pipe.

launcher ── daemonize() ──► fork → setsid → fork ──► grandchild = your daemon
   ▲                                                     │
   └───────────── readiness / error  (pipe) ─────────────┘

That byte releases the launcher, which exits 0. If the daemon instead dies or calls report_error() before notifying, the launcher exits non-zero with the matching sysexits.h code -- so it sees a real result, not a process that detached and then crashed.

Steps 1-2 must run single-threaded: spawn threads, async runtimes, or accept loops only after drop_privileges() returns. See Safety for why.

Install

cargo install blivet   # the `daemonize` CLI -- verify with `daemonize --version`
cargo add blivet        # the library

The crate is blivet; the installed binary is daemonize. See Library for the API, or CLI for command-line use.

Library

The minimal example above writes a pidfile and signals readiness. A fuller setup adds a lock file, log redirection, and a working directory. As with daemonize(1), the defaults are standard daemon behavior -- stdout/stderr go to /dev/null and the working directory becomes / -- so use absolute paths and redirect any output you want to keep:

use blivet::{daemonize, DaemonConfig};

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let mut config = DaemonConfig::new();
    config
        .pidfile("/var/run/myapp.pid")
        .lockfile("/var/run/myapp.lock")
        .stdout("/var/log/myapp.out")
        .stderr("/var/log/myapp.err")
        .chdir("/var/lib/myapp");

    let mut ctx = daemonize(&config)?;
    // fallible initialization here (open sockets, load config, …)
    ctx.notify_parent()?; // signal readiness; the launcher exits 0
    Ok(())
}

Split-phase privilege dropping. Do privileged work (bind port 80, chroot, set rlimits) between daemonize() and drop_privileges():

use blivet::{daemonize, DaemonConfig};

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let mut config = DaemonConfig::new();
    config.pidfile("/var/run/myapp.pid").user("www-data").group("www-data");

    let mut ctx = daemonize(&config)?;

    // Still root here -- bind a privileged port, chroot, set rlimits…
    // let listener = TcpListener::bind("0.0.0.0:80")?;

    ctx.chown_paths()?;     // transfer file ownership to the target user
    ctx.drop_privileges()?; // then switch user/group
    ctx.notify_parent()?;
    // now running as www-data…
    Ok(())
}

Foreground mode. For systemd, containers, or debugging: skip forking while applying all other setup (umask, chdir, signal reset, …). Stdout/stderr stay inherited unless explicitly redirected with .stdout()/.stderr():

config.foreground(true).close_fds(false); // keep supervisor-passed fds

See examples/echo_server.rs for a complete, runnable daemonized TCP echo server with signal-based shutdown and pidfile cleanup.

Entry points

• daemonize(&config) -- the safe, recommended entry point. It verifies the process is single-threaded, then daemonizes. Available on Linux, macOS, FreeBSD, NetBSD, and OpenBSD (it reads the kernel thread count: /proc/self/status on Linux, proc_pidinfo on macOS, sysctl on the BSDs). On any other target it is a #[deprecated] stub that panics; use the unchecked form below.
• unsafe { daemonize_unchecked(&config) } -- the escape hatch, on all Unix platforms. It skips the thread-count check, so you must guarantee the process is single-threaded.

DaemonContext::drop_privileges() mirrors this split: it is safe and checked (panicking if a user switch is configured while multithreaded), with unsafe { drop_privileges_unchecked() } as the opt-out. See Safety.

Safety

Forking a multithreaded process is unsound: mutexes held by other threads stay locked forever in the child, deadlocking it. A second thread-unsafe step follows -- drop_privileges() calls setenv (USER/HOME/LOGNAME) when switching users -- so the single-threaded window runs from the fork through drop_privileges() (see How it works). Spawn threads, an async runtime, or an accept loop only after drop_privileges() returns -- or after daemonize() returns if you don't switch users.

Both checked entry points read the kernel thread count and panic if violated: daemonize() at the fork, and drop_privileges() at its setenv (only when a user switch is configured -- the sole setenv path; a group-only switch is not guarded). daemonize_unchecked() and drop_privileges_unchecked() are the unsafe opt-outs for callers who manage the contract themselves, or who run on a target without a thread-count source.

API reference

Full reference is on docs.rs; this is the shape of it.

DaemonConfig -- a builder of infallible &mut self setters; validation is deferred to validate(), which daemonize() runs for you. Settings: pidfile, lockfile, stdout/stderr (+ append), chdir, umask, user/group, foreground, close_fds, cleanup_on_drop, and env. Defaults worth knowing: working directory /, stdout/stderr /dev/null, and close_fds and cleanup_on_drop both true.

DaemonContext -- returned by a successful daemonize(); owns the lockfile and notification pipe. The methods you reach for most: notify_parent(), chown_paths() / drop_privileges(), cleanup() / cleanup_on_term_signals(), and report_error() / report_error_msg() / notify_parent_or_report().

Errors & exit codes

DaemonizeError has sixteen variants covering validation, fork, setsid, lock, permission, chown, exec, and parent-notify failures, plus a caller-supplied Application variant. Each maps to a sysexits.h exit code via exit_code(), so failures reach the shell with a meaningful status:

Recipes

Pidfile cleanup on signals

When cleanup_on_drop is true (the default), the pidfile is removed when DaemonContext is dropped. But Drop does not run when the process is killed by a signal (SIGTERM, SIGKILL, …) -- which is how most daemons stop -- so the pidfile would be left behind. The built-in fix installs async-signal-safe handlers that remove the pidfile on SIGINT/SIGTERM, then re-raise so the process still terminates normally:

use blivet::{daemonize, DaemonConfig};

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let mut config = DaemonConfig::new();
    config.pidfile("/var/run/myapp.pid");
    let mut ctx = daemonize(&config)?;

    ctx.cleanup_on_term_signals()?; // or cleanup_on_signals(&[...]) for custom signals
    ctx.notify_parent()?;
    // … daemon work …
    Ok(())
}

Library-only. The daemonize CLI cannot do this: it execs the target program, and exec resets custom signal handlers to their default disposition, so a CLI-launched program must clean up its own pidfile.

If you already run your own signal loop (e.g. for graceful shutdown), you don't need the built-in handler: let the loop exit, then call cleanup() -- or just let ctx drop. The signal_hook crate is one way to drive that loop; blivet does not re-export it, so cargo add signal_hook first:

use std::sync::atomic::{AtomicBool, Ordering};
use std::sync::Arc;
use blivet::{daemonize, DaemonConfig};

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let mut config = DaemonConfig::new();
    config.pidfile("/var/run/myapp.pid");
    let mut ctx = daemonize(&config)?;
    ctx.notify_parent()?;

    let running = Arc::new(AtomicBool::new(true));
    signal_hook::flag::register(signal_hook::consts::SIGTERM, Arc::clone(&running))?;
    signal_hook::flag::register(signal_hook::consts::SIGINT, Arc::clone(&running))?;

    while running.load(Ordering::Relaxed) {
        // … daemon work …
    }
    ctx.cleanup(); // or just let ctx drop
    Ok(())
}

Reporting your own failures

If startup work in the privileged init window fails (a socket bind, a database connect), report it to the launcher with a sysexits.h code of your choosing via report_error_msg -- no need to construct a DaemonizeError by hand:

let listener = match TcpListener::bind("0.0.0.0:80") {
    Ok(l) => l,
    // 71 == EX_OSERR; the launcher prints the message and exits with this code.
    Err(e) => ctx.report_error_msg(71, format!("bind failed: {e}")),
};

Propagating exit codes

The sysexits.h codes only reach the shell if you use them. The idiomatic fn main() -> Result<(), E> prints the error via Termination and exits 1, ignoring exit_code(). To preserve the codes, drive exit_code() yourself:

use blivet::{daemonize, DaemonConfig, DaemonizeError};

fn main() {
    if let Err(e) = run() {
        eprintln!("{e}");
        std::process::exit(e.exit_code() as i32); // e.g. 77 for PermissionDenied
    }
}

fn run() -> Result<(), DaemonizeError> {
    let config = DaemonConfig::new();
    let mut ctx = daemonize(&config)?;
    // … application init …
    // notify_parent() returns DaemonizeError (NotifyFailed, exit 71), so `?`
    // keeps a single error type and preserves the exit code.
    ctx.notify_parent()?;
    Ok(())
}

CLI

The daemonize binary wraps any program as a daemon, applying the same setup as the library:

# Simplest: daemonize a program
daemonize -- /usr/bin/my-server --port 8080

# Typical service: pidfile, log redirect, working dir, drop to an unprivileged user
daemonize -p /var/run/myapp.pid -o /var/log/myapp.log -c /var/lib/myapp \
  -u www-data -g www-data -- /usr/bin/my-server

The launcher blocks until the daemon successfully execs, then exits 0. On failure (lockfile conflict, permission denied, exec error) it prints the error to stderr and exits with a sysexits.h code -- the same codes the library returns (see Errors & exit codes). When -u/-g are given, the CLI transfers ownership of the pidfile, lockfile, and log files to the target user/group before dropping privileges, so the daemon can keep writing to them.

Minimum supported Rust version

1.85

Why the name blivet?

A blivet is the "impossible pitchfork" optical illusion, also known as the devil's fork, where the prongs are mysteriously detached from the base. Daemons are created by forking to detach from their parent terminal.

License

Licensed under either of Apache License, Version 2.0 or MIT License at your option.