heddle-cli 0.2.3

// SPDX-License-Identifier: Apache-2.0
//! `heddle daemon …` CLI dispatchers.
//!
//! Three verbs:
//!
//! * `serve` — runs the foreground daemon. Linux + `--features mount`
//!   only; everywhere else it returns the standard
//!   `virtualized_unsupported_error`.
//! * `status` — sends `health` to a running daemon and prints the
//!   reply. No-op success when the daemon isn't running, so
//!   operators can run `heddle daemon status` as a probe.
//! * `stop` — sends `shutdown`, waits for the endpoint file to
//!   disappear *and* the daemon PID to die, then sweeps any
//!   leftover mounts as a safety net. The combined wait gives
//!   callers a hard post-condition (see `cmd_daemon_stop`).

use std::time::Duration;

use anyhow::{Result, anyhow};
use repo::daemon::{
    MountDaemonRequest, MountDaemonResponse, load_endpoint, mount_daemon_endpoint_path, pid_alive,
};

use super::client::{rpc, sweep_stale_mounts};
use crate::cli::Cli;

#[cfg(all(target_os = "linux", feature = "mount"))]
pub fn cmd_daemon_serve(cli: &Cli) -> Result<()> {
    let repo_root = resolve_repo_root(cli)?;
    super::server::run_mount_daemon(&repo_root)
}

#[cfg(not(all(target_os = "linux", feature = "mount")))]
pub fn cmd_daemon_serve(_cli: &Cli) -> Result<()> {
    Err(
        crate::cli::commands::mount_lifecycle::virtualized_unsupported_error()
            .context("heddle daemon serve"),
    )
}

pub fn cmd_daemon_status(cli: &Cli) -> Result<()> {
    let repo_root = resolve_repo_root(cli)?;
    let response = rpc(&repo_root, &MountDaemonRequest::Health {}, false)?;
    match response {
        Some(MountDaemonResponse::Health {
            version,
            ok,
            uptime_s,
            mount_count,
        }) => {
            println!(
                "daemon: ok={ok} version={version} uptime_s={uptime_s} mount_count={mount_count}"
            );
            Ok(())
        }
        Some(MountDaemonResponse::Error { code, message, .. }) => {
            Err(anyhow!("daemon health failed: [{code}] {message}"))
        }
        Some(other) => Err(anyhow!("unexpected daemon response: {other:?}")),
        None => {
            println!(
                "daemon: not running (no live endpoint at {})",
                mount_daemon_endpoint_path(&repo_root).display()
            );
            Ok(())
        }
    }
}

/// Post-condition contract for `cmd_daemon_stop`: when this returns
/// `Ok(())` after a live-daemon shutdown, the caller may rely on
/// **all four** of the following being true:
///
/// 1. The daemon process (whose PID was advertised in the endpoint
///    file) has exited (`kill -0` returns `ESRCH`).
/// 2. `<repo>/.heddle/state/heddled.endpoint.json` no longer exists.
/// 3. `<repo>/.heddle/state/mounts.json` no longer exists. The
///    daemon's `MountRegistry::shutdown_all` removes it before
///    `remove_endpoint`, and the CLI-side `sweep_stale_mounts` runs
///    as a safety-net (idempotent — both use `fs::remove_file` and
///    swallow `NotFound`).
/// 4. Any FUSE mountpoints the daemon owned are unmounted (best-effort
///    via the `BackgroundSession` drop in `LiveMount::shutdown`, with
///    `fusermount -u` as a fallback inside `sweep_stale_mounts`).
///
/// Two timeouts are layered to make the contract observable rather
/// than hopeful: 2 s for the endpoint file to disappear (proof the
/// daemon's `run_mount_daemon` reached its post-shutdown cleanup), and
/// a further 2 s for the PID to be reaped. Either can elapse without
/// failing the call — the safety-net sweep still runs — but together
/// they make the integration-test assertions deterministic.
pub fn cmd_daemon_stop(cli: &Cli) -> Result<()> {
    let repo_root = resolve_repo_root(cli)?;
    let endpoint_path = mount_daemon_endpoint_path(&repo_root);
    // Capture the daemon PID *before* sending shutdown so we can
    // probe it via `kill -0` after the endpoint file is gone. If the
    // endpoint file has no recorded PID (v1-era files, or a future
    // schema change) we just skip the PID wait — the endpoint-gone
    // observation is still load-bearing.
    let recorded_pid = load_endpoint(&endpoint_path).ok().and_then(|e| e.pid);
    match rpc(&repo_root, &MountDaemonRequest::Shutdown {}, false)? {
        Some(MountDaemonResponse::Shutdown { ok: true, .. }) => {}
        Some(MountDaemonResponse::Error { code, message, .. }) => {
            return Err(anyhow!("daemon refused shutdown: [{code}] {message}"));
        }
        Some(other) => return Err(anyhow!("unexpected daemon response: {other:?}")),
        None => {
            println!("daemon: not running");
            return Ok(());
        }
    }
    // Phase 1: wait up to 2 s for the endpoint file to disappear.
    // The daemon's `run_mount_daemon` removes it *after*
    // `MountRegistry::shutdown_all` (which removes `mounts.json`),
    // so endpoint-gone implies mounts.json-gone on the daemon side.
    for _ in 0..40 {
        if !endpoint_path.exists() {
            break;
        }
        std::thread::sleep(Duration::from_millis(50));
    }
    // Phase 2: wait up to a further 2 s for the daemon process
    // itself to exit. Without this, the endpoint-gone observation
    // races the daemon's final `info!("heddle daemon exiting")` +
    // process teardown — a caller probing PID liveness right after
    // `daemon stop` returns could still see the PID briefly. Polling
    // here turns the post-condition from "shutdown is in flight"
    // into "shutdown is complete".
    if let Some(pid) = recorded_pid {
        for _ in 0..40 {
            if !pid_alive(pid) {
                break;
            }
            std::thread::sleep(Duration::from_millis(50));
        }
    }
    // Sweep any leftover registry entries as a last-resort safety
    // net for crash-during-shutdown scenarios. Idempotent: in the
    // happy path the daemon has already removed `mounts.json`, so
    // this is a no-op.
    sweep_stale_mounts(&repo_root);
    println!("daemon: stopped");
    Ok(())
}

fn resolve_repo_root(cli: &Cli) -> Result<std::path::PathBuf> {
    if let Some(root) = cli.repo.as_ref() {
        return Ok(root.clone());
    }
    let repo = repo::Repository::open(&std::env::current_dir()?)?;
    Ok(repo.root().to_path_buf())
}