mj_controller/daemon/

process.rs

use super::diagnostics::ServingPhase;
use super::*;
use std::time::Instant;

/// PID of the process this daemon must not outlive, if one was requested.
///
/// Tests start daemons that no client ever attaches to, so idle exit cannot
/// retire them, and a test process that dies without unwinding never runs its
/// teardown. Naming an owner makes the daemon responsible for its own lifetime.
pub(super) fn owner_pid_to_watch() -> Result<Option<u32>> {
    let Some(value) = mj_core::config::env_override("DAEMON_OWNER_PID") else {
        return Ok(None);
    };
    let pid: u32 = value
        .trim()
        .parse()
        .map_err(|_| anyhow!("MJ_DAEMON_OWNER_PID must be a process id, but it is {value:?}"))?;
    ensure!(
        process_is_alive(pid),
        "MJ_DAEMON_OWNER_PID names process {pid}, which is not running"
    );
    Ok(Some(pid))
}

pub async fn run_daemon_process() -> Result<()> {
    // Checked before the store is locked so a bad value fails fast and leaves
    // no daemon state behind.
    let owner_pid = owner_pid_to_watch()?;
    let opening = Instant::now();
    let guard = ControllerStoreGuard::acquire()?;
    let database_writer = guard.start_database_writer()?;
    log_startup_phase("open the store", opening);
    let epilogue_started = AtomicBool::new(false);
    let mut outcome = run_daemon_runtime(&epilogue_started, owner_pid).await;
    if !epilogue_started.load(Ordering::Acquire) {
        // Initialization failed before the runtime-owned epilogue existed.
        // The same process-level bound still applies to closing the writer.
        spawn_shutdown_watchdog();
    }
    // The writer commits every write it accepted before it stops, so this is
    // the one wait after the epilogue that durability requires. Timed, so a
    // slow exit says whether the writer or the runtime held it.
    let writer_started = Instant::now();
    let writer_shutdown = tokio::task::spawn_blocking(move || database_writer.shutdown())
        .await
        .context("database writer shutdown task panicked")
        .and_then(std::convert::identity);
    let writer_took = writer_started.elapsed();
    if writer_took >= Duration::from_millis(250) {
        tracing::info!(
            duration_ms = writer_took.as_millis(),
            "database writer shut down"
        );
    }
    record_daemon_cleanup(&mut outcome, "shut down database writer", writer_shutdown);
    outcome
}

pub(super) async fn run_daemon_runtime(
    epilogue_started: &AtomicBool,
    owner_pid: Option<u32>,
) -> Result<()> {
    let progress = super::diagnostics::DaemonProgressMonitor::start()?;
    // Before this daemon can start a checkpoint, so every checkpoint file older
    // than this belongs to a checkpoint that no longer runs.
    let started = SystemTime::now();
    let startup_work = crate::upgrade::activity("daemon startup recovery")?;
    // Freeze worker sources before any session can be created or upgraded.
    // Copying binaries belongs on a blocking task, never the runtime event loop.
    tokio::task::spawn_blocking(crate::controller::pin_worker_binary_sources)
        .await
        .context("worker source snapshot task failed")??;
    // Lock files outlive the SSH masters they guarded (launch finding R3-11).
    tokio::task::spawn_blocking(crate::targets::SshSessions::remove_stale_master_locks)
        .await
        .context("SSH master lock cleanup task failed")?;
    Controller::recover_config_id_rename()?;
    let config = Config::load()?;
    crate::database::recover_interrupted_checkpointing_sessions(
        &chrono::Utc::now().to_rfc3339_opts(chrono::SecondsFormat::Secs, true),
    )?;
    crate::controller::reconcile_managed_checkpoint_archives()?;

    let loading = Instant::now();
    let controller = tokio::task::spawn_blocking(|| {
        let mut controller = Controller::load()?;
        controller.prepare_persisted_sessions()?;
        Ok::<_, anyhow::Error>(controller)
    })
    .await
    .context("prepare persisted daemon sessions")??;
    log_startup_phase("load persisted sessions", loading);
    // A local session an earlier release started from a profile home keeps
    // running from it until it is next staged. The link has to be in place
    // before any launch configuration is refreshed or any credential sync runs.
    {
        let state = controller.state.clone();
        tokio::task::spawn_blocking(move || {
            crate::controller::local_profile_homes::link_profile_homes_of_earlier_sessions(&state)
        })
        .await
        .context("earlier sessions' profile home link task failed")?;
    }
    let listener = TcpListener::bind(SocketAddr::new(IpAddr::V4(Ipv4Addr::LOCALHOST), 0))
        .await
        .context("bind Mjolnir daemon loopback endpoint")?;
    let metadata = DaemonMetadata {
        protocol_version: PROTOCOL_VERSION,
        pid: std::process::id(),
        address: listener.local_addr()?,
        token: random_hex::<32>()?,
        started_at: chrono::Utc::now().to_rfc3339_opts(chrono::SecondsFormat::Secs, true),
        // The version, with the revision and build times as semver build
        // metadata that older clients ignore; see `mj_client::build_identity`.
        build_version: mj_client::build_identity::this_build().published(),
    };
    let workspaces = tokio::task::spawn_blocking(crate::database::list_workspaces)
        .await
        .context("daemon workspace load task panicked")??;
    let mut remote = if config.phone.enabled {
        Some(spawn_remote_session_manager()?)
    } else {
        None
    };

    // Start the primary manager after store preparation. Bootstrap failures
    // below have an awaited epilogue just like failures in the serving loop.
    let (delegation_tx, delegation_updates) = crate::session_manager::delegation_channel();
    let manager = crate::session_manager::spawn_session_manager_observed(Some(delegation_tx))?;
    let manager_targets = manager.targets;
    manager_targets.send_replace(dashboard_worker_targets(&controller));
    let manager_updates = manager.updates;
    let manager_control = manager.control.clone();
    let manager_shutdown = manager.shutdown;
    let mut recovery = crate::recovery::RecoveryCoordinator::spawn(manager_control.clone());
    let recovery_observer = recovery.observer();
    // Shares the recovery gate, so a recovery copy and a worker upgrade never
    // act on one session at the same time.
    let mut worker_upgrades = crate::worker_upgrade::WorkerUpgradeCoordinator::spawn(
        manager_control.clone(),
        &recovery_observer,
    );
    let state = Arc::new(RuntimeState::new(
        manager_control.clone(),
        Controller {
            config: controller.config.clone(),
            state: controller.state.clone(),
        },
        recovery_observer.clone(),
        worker_upgrades.observer(),
        workspaces,
    ));
    // Views are published only from the serving loop below, so the first
    // targets are installed before any can arrive.
    state.owner().install_relay_sessions(
        manager_targets
            .borrow()
            .iter()
            .map(|target| target.session_id.clone())
            .collect(),
    );
    let cancellation = crate::termination::Coordinator::install().token();
    // Bootstrap already owns live managers. Capture its error so those owners
    // are shut down before the process-level writer can be closed.
    let bootstrapping = Instant::now();
    let bootstrap = async {
        let move_operations = blocking(crate::database::load_move_operations).await?;
        let move_sessions = move_operations
            .iter()
            .map(|operation| operation.selection.session_id.clone())
            .collect::<BTreeSet<_>>();
        let mut lifecycle_owned = state.recover_moves(move_operations)?;
        let restart_intents = blocking(crate::database::load_session_restarts).await?;
        lifecycle_owned.extend(
            restart_intents
                .iter()
                .map(|intent| intent.session_id.clone()),
        );
        lifecycle_owned.extend(state.recover_restarts(restart_intents)?);
        lifecycle_owned.extend(move_sessions.iter().cloned());
        state.resume_retained_cleanups();
        state.resume_startup_cleanups(true);
        state.restore_startup_deliveries(&cancellation).await?;
        Ok::<_, anyhow::Error>((move_sessions, lifecycle_owned))
    }
    .await;
    log_startup_phase("recover moves and startup deliveries", bootstrapping);
    let (move_sessions, move_owned) = match bootstrap {
        Ok(ownership) => ownership,
        Err(error) => {
            epilogue_started.store(true, Ordering::Release);
            spawn_shutdown_watchdog();
            cancellation.cancel();
            let mut outcome = Err(error);
            record_daemon_cleanup(
                &mut outcome,
                "shut down bootstrap worker upgrade coordinator",
                worker_upgrades.shutdown().await,
            );
            record_daemon_cleanup(
                &mut outcome,
                "shut down bootstrap recovery coordinator",
                recovery.shutdown().await,
            );
            record_daemon_cleanup(
                &mut outcome,
                "shut down turn review host after bootstrap failure",
                state
                    .review_host()
                    .shutdown()
                    .await
                    .map_err(anyhow::Error::msg),
            );
            record_daemon_cleanup(
                &mut outcome,
                "cancel bootstrap lifecycle operations",
                state.cancel_and_wait_lifecycles().await,
            );
            record_daemon_cleanup(
                &mut outcome,
                "drain bootstrap startup deliveries",
                state.cancel_and_join_startup_prompts().await,
            );
            if let Some(remote) = remote.take() {
                record_daemon_cleanup(
                    &mut outcome,
                    "shut down bootstrap remote session manager",
                    remote.shutdown.shutdown().await,
                );
            }
            record_daemon_cleanup(
                &mut outcome,
                "shut down bootstrap session manager",
                manager_shutdown.shutdown().await,
            );
            return outcome;
        }
    };
    // Checkpoint files that a cancelled or interrupted checkpoint left in a
    // local worker root. Deleting them can take minutes after a long leak, so
    // this runs beside the daemon rather than before it serves. It stops at
    // shutdown, and the next start finishes it.
    let checkpoint_sweep = {
        let cancellation = cancellation.clone();
        tokio::task::spawn_blocking(move || {
            crate::controller::sweep_local_checkpoint_leftovers(
                &crate::targets::BoundedProcessExecutor::new(Duration::from_secs(15)),
                started,
                &|| cancellation.is_cancelled(),
            );
        })
    };
    let mut update_feed = continuation::spawn(state.clone(), manager_updates, cancellation.clone());

    let (delegation_services, mut delegation_task) = super::delegation::spawn(
        state.clone(),
        manager_control.clone(),
        delegation_updates,
        cancellation.clone(),
    );

    // CPU publication is bounded and reconstructed from workers. It holds no
    // admission during handoff and exits with the daemon's cancellation token.
    let cpu_publication = {
        let mut receiver = manager.session_cpu;
        let state = state.clone();
        let cancellation = cancellation.clone();
        tokio::spawn(async move {
            let mut next_publication = tokio::time::Instant::now();
            loop {
                tokio::select! {
                    _ = cancellation.cancelled() => return,
                    changed = receiver.changed() => {
                        if changed.is_err() {
                            tracing::error!("session CPU publication channel closed");
                            return;
                        }
                    }
                }
                // Publish the first update promptly, then coalesce other actors'
                // updates. A separate polling phase adds a needless ten seconds.
                tokio::select! {
                    _ = cancellation.cancelled() => return,
                    _ = tokio::time::sleep_until(next_publication) => {}
                }
                receiver.borrow_and_update();
                state.publish_revision();
                next_publication = tokio::time::Instant::now() + Duration::from_secs(10);
            }
        })
    };
    let project_catalog = tokio::spawn(state.projects().run(cancellation.child_token()));
    // Before the web viewer starts: it shows this service's readings.
    let capacity_service =
        super::capacity::spawn_capacity_service(state.clone(), cancellation.clone());

    let target_refresh = spawn_manager_target_refresher(
        manager_targets.clone(),
        cancellation.clone(),
        state.clone(),
    );
    let image_refresh = spawn_image_refresher(
        {
            let state = state.clone();
            move || state.with_config(crate::controller::image_refresh_plan)
        },
        {
            // A background download is the daemon's own work, not a session's,
            // so these notices carry an empty session id and reach every
            // workspace.
            let state = state.clone();
            move |report| {
                let text = match report {
                    crate::pollers::ImageRefreshReport::Started { host, image } => {
                        format!("Downloading image {image} for {host}\u{2026}")
                    }
                    crate::pollers::ImageRefreshReport::Pulled { host, image } => {
                        format!("Image {image} is ready on {host}.")
                    }
                    crate::pollers::ImageRefreshReport::Failed { host, image, error } => {
                        format!("Could not pull image {image} on {host}: {error}")
                    }
                };
                state.push_notice("", text);
            }
        },
        cancellation.clone(),
    );
    let exit_when_idle = mj_core::config::env_override_os("DAEMON_EXIT_WHEN_IDLE").is_some();
    let mut idle_tick = tokio::time::interval(Duration::from_millis(100));
    idle_tick.set_missed_tick_behavior(tokio::time::MissedTickBehavior::Delay);
    let mut owner_tick = tokio::time::interval(Duration::from_millis(500));
    owner_tick.set_missed_tick_behavior(tokio::time::MissedTickBehavior::Delay);
    let mut recovery_tick = tokio::time::interval(Duration::from_millis(250));
    recovery_tick.set_missed_tick_behavior(tokio::time::MissedTickBehavior::Delay);
    let mut background_policy_tick = tokio::time::interval(Duration::from_secs(1));
    background_policy_tick.set_missed_tick_behavior(tokio::time::MissedTickBehavior::Delay);
    // How often the harness readiness wait is looked at. The wait itself is
    // minutes long, so this only bounds how late a failure is noticed, and it
    // reads in-memory state rather than the store.
    let mut readiness_tick = tokio::time::interval(Duration::from_secs(5));
    readiness_tick.set_missed_tick_behavior(tokio::time::MissedTickBehavior::Delay);
    let mut progress_tick = tokio::time::interval(Duration::from_secs(1));
    progress_tick.set_missed_tick_behavior(tokio::time::MissedTickBehavior::Delay);
    let (interrupted_close_tx, mut interrupted_close_rx) = tokio::sync::mpsc::unbounded_channel();
    let mut interrupted_close_tasks = Vec::new();
    for session_id in interrupted_suspend_session_ids(&controller) {
        if move_owned.contains(&session_id) {
            continue;
        }
        let recovery_state = state.clone();
        let recovery_shutdown = cancellation.clone();
        let updates = interrupted_close_tx.clone();
        let upgrade_work = startup_work.clone();
        let interrupted_close_task = tokio::spawn(async move {
            let _upgrade_work = upgrade_work;
            let result = tokio::select! {
                result = recovery_state.suspend_session(session_id.clone()) => result,
                () = recovery_shutdown.cancelled() => return,
            }
            .map(|()| crate::pollers::LifecycleSuccess::Closed)
            .map_err(|error| format!("{error:#}"));
            if updates
                .send(crate::pollers::LifecycleUpdate {
                    session_id,
                    result,
                    deferred_cleanup: false,
                })
                .is_err()
            {
                tracing::debug!("suspension recovery receiver stopped");
            }
        });
        interrupted_close_tasks.push(interrupted_close_task);
    }
    // A destroy that failed left its record in `Error` with the target still
    // present; it is finished the way `mj destroy` finishes it. The branch is
    // kept, because whether to delete it was that command's choice and is not
    // recorded.
    for session_id in interrupted_destroy_session_ids(&controller) {
        let recovery_state = state.clone();
        let recovery_shutdown = cancellation.clone();
        let updates = interrupted_close_tx.clone();
        let upgrade_work = startup_work.clone();
        interrupted_close_tasks.push(tokio::spawn(async move {
            let _upgrade_work = upgrade_work;
            let result = tokio::select! {
                result = recovery_state.force_destroy_session(
                    session_id.clone(),
                    crate::daemon::BranchDisposition::Keep,
                ) => result,
                () = recovery_shutdown.cancelled() => return,
            }
            .map(|()| crate::pollers::LifecycleSuccess::ForceDestroyed)
            .map_err(|error| format!("{error:#}"));
            let _ = updates.send(crate::pollers::LifecycleUpdate {
                session_id,
                result,
                deferred_cleanup: false,
            });
        }));
    }
    // Whatever is still in an in-flight lifecycle state now has no owner: the
    // moves, the interrupted closes, and the checkpointing rows above are
    // every operation that legitimately resumes. The list comes from the
    // startup snapshot, so a session created after this cannot be caught by
    // it, and the writes run off this path because they touch the database.
    let reconciliation = {
        let unowned = unowned_interrupted_lifecycles(
            &controller,
            &move_owned.union(&move_sessions).cloned().collect(),
        );
        (!unowned.is_empty()).then(|| {
            let state = state.clone();
            let upgrade_work = startup_work.clone();
            tokio::spawn(async move {
                let _upgrade_work = upgrade_work;
                let reconciled = tokio::task::spawn_blocking(move || {
                    let mut controller = Controller::load()?;
                    let mut reconciled = 0usize;
                    for (session_id, cause) in unowned {
                        match controller.fail_interrupted_lifecycle(&session_id, &cause) {
                            Ok(true) => {
                                tracing::warn!(%session_id, %cause, "session left in flight by a daemon restart marked failed");
                                reconciled += 1;
                            }
                            Ok(false) => {}
                            Err(error) => tracing::warn!(
                                %session_id,
                                error = format!("{error:#}"),
                                "could not reconcile an interrupted lifecycle state"
                            ),
                        }
                    }
                    anyhow::Ok(reconciled)
                })
                .await;
                match reconciled {
                    Ok(Ok(0)) => {}
                    Ok(Ok(_)) => refresh_runtime_controller(&state).await,
                    Ok(Err(error)) => tracing::warn!(
                        error = format!("{error:#}"),
                        "could not load the controller to reconcile interrupted lifecycles"
                    ),
                    Err(error) => {
                        tracing::warn!(%error, "interrupted lifecycle reconciliation task failed");
                    }
                }
            })
        })
    };
    // Rows left as tombstones by an older build, or by a discard that could
    // not finish before the daemon stopped. The list comes from the startup
    // snapshot, so a session that becomes lost after this is handled by the
    // view watcher instead.
    let tombstone_sweep = {
        let tombstones = tombstone_session_ids(&controller);
        (!tombstones.is_empty()).then(|| {
            let state = state.clone();
            let upgrade_work = startup_work.clone();
            tokio::spawn(async move {
                let _upgrade_work = upgrade_work;
                for session_id in tombstones {
                    state.discard_lost_session(session_id).await;
                }
            })
        })
    };
    let mut phone_publisher: Option<RemoteSessionPublisher> = None;
    let mut phone_targets = None;
    let mut prepared_targets = manager_targets.subscribe();
    let mut phone_task = None;
    let mut remote_request_bridge = None;
    if let Some(remote) = remote.take() {
        remote
            .targets
            .send_replace(prepared_targets.borrow_and_update().clone());
        phone_targets = Some(remote.targets.clone());
        phone_publisher = Some(remote.publisher.clone());
        remote_request_bridge = Some(spawn_remote_request_bridge(
            remote.requests,
            manager_control.clone(),
        ));
        phone_task = Some(spawn_phone_server(
            config.phone,
            cancellation.clone(),
            state.clone(),
            SessionManagerChannels {
                session_cpu: remote.control.session_cpu.clone(),
                targets: remote.targets,
                control: remote.control,
                updates: remote.updates,
                shutdown: remote.shutdown,
            },
            delegation_services,
        ));
    } else {
        state.set_phone_status(WebViewerStatus::Disabled);
        state.web_viewer.publish(crate::server::WebViewerAccess::Unavailable("Web access is disabled. Enable [phone].enabled in your configuration, then restart the daemon.".into()));
    }
    let daemon_metadata_path = metadata_path();
    let mut client_tasks = tokio::task::JoinSet::new();
    let mut delegation_outcome = None;
    let mut cache_task = tokio::spawn(crate::controller::mbx::service::run(
        state.clone(),
        cancellation.clone(),
    ));
    let mut mailbox_task = tokio::spawn(crate::mailbox_outbox::run(
        state.clone(),
        manager_control.clone(),
        cancellation.clone(),
    ));
    let mut github_watch_task = tokio::spawn(crate::github_watch::service::run(
        state.clone(),
        cancellation.clone(),
    ));

    let mut cache_outcome = None;
    let mut mailbox_outcome = None;
    let mut github_watch_outcome = None;

    // Everything a client can use is initialized before this atomic
    // publication. From here on every exit, including an error from the test
    // hook or the event loop, flows through the same bounded epilogue.
    let mut outcome = async {
        write_metadata(&daemon_metadata_path, &metadata)?;
        reach_test_hook("daemon_metadata_before_listening").await?;
        tracing::info!("the daemon is serving");
        drop(startup_work);
        loop {
            progress.serving_tick();
            tokio::select! {
                _ = cancellation.cancelled() => break,
                _ = progress_tick.tick() => {},
                changed = prepared_targets.changed(), if phone_targets.is_some() => {
                    progress.phase(ServingPhase::Targets);
                    changed.context("daemon worker target publication stopped")?;
                    if let Some(targets) = &phone_targets {
                        targets.send_replace(prepared_targets.borrow_and_update().clone());
                    }
                }
                result = &mut cache_task => {
                    cache_outcome = Some(result.context("machine cache service failed").and_then(|result| result));
                    anyhow::bail!("machine build cache service stopped unexpectedly");
                }
                result = &mut mailbox_task => {
                    mailbox_outcome = Some(result.context("mailbox outbox service failed").and_then(|result| result));
                    anyhow::bail!("mailbox outbox service stopped unexpectedly");
                }
                result = &mut github_watch_task => {
                    github_watch_outcome = Some(result.context("GitHub watch service failed").and_then(|result| result));
                    anyhow::bail!("GitHub watch service stopped unexpectedly");
                }
                result = &mut delegation_task => {
                    delegation_outcome = Some(result.map_err(anyhow::Error::from).and_then(|result| result));
                    anyhow::bail!("delegation coordinator stopped unexpectedly");
                }
                _ = idle_tick.tick(), if exit_when_idle && state.ever_attached.load(Ordering::Acquire) => {
                    progress.phase(ServingPhase::IdleCheck);
                    state.prune_dead_clients();
                    if state.attachments().is_empty() {
                        break;
                    }
                }
                _ = owner_tick.tick(), if owner_pid.is_some() => {
                    progress.phase(ServingPhase::OwnerCheck);
                    if let Some(owner) = owner_pid
                        && !process_is_alive(owner)
                    {
                        tracing::info!(owner_pid = owner, "daemon owner process exited; shutting down");
                        break;
                    }
                }
                _ = recovery_tick.tick() => {
                    progress.phase(ServingPhase::Recovery);
                    while let Some(result) = recovery.try_result() {
                        if let Err(error) = &result.outcome {
                            // A deferred copy found the agent working. That is
                            // the normal state of a session in use, so it is
                            // news, not a fault.
                            if result.deferred {
                                tracing::info!(session_id = %result.session_id, %error, "recovery copy deferred: agent is working");
                            } else if result.cancelled {
                                // The daemon asked for this cancel (a suspend
                                // or close superseded the copy), so it is not
                                // a failure of the checkpoint.
                                tracing::info!(session_id = %result.session_id, %error, reason = "superseded by suspend", "daemon recovery checkpoint cancelled");
                            } else {
                                tracing::warn!(session_id = %result.session_id, %error, "daemon recovery checkpoint failed");
                            }
                        }
                        refresh_runtime_controller(&state).await;
                    }
                    while let Some(result) = worker_upgrades.try_result() {
                        report_worker_upgrade(&state, &result).await;
                    }
                }
                _ = background_policy_tick.tick() => {
                    progress.phase(ServingPhase::BackgroundPolicy);
                    state.refresh_background_policies();
                    state.resume_startup_cleanups(false);
                }
                _ = readiness_tick.tick() => {
                    progress.phase(ServingPhase::Readiness);
                    // A live session whose harness never advertised itself
                    // takes no prompt and reports no failure, so nothing else
                    // ever ends its wait (#1090). The store write runs off
                    // this loop.
                    for unready in state.sessions_without_a_usable_harness() {
                        let state = state.clone();
                        client_tasks.spawn(async move {
                            state.fail_unready_session(unready).await;
                        });
                    }
                }
                completed = interrupted_close_rx.recv() => {
                    progress.phase(ServingPhase::SuspensionRecovery);
                    if let Some(completed) = completed {
                        let recovered = completed.result.is_ok();
                        if let Err(error) = completed.result {
                            tracing::warn!(session_id = %completed.session_id, %error, "daemon could not resume interrupted close");
                        }
                        refresh_runtime_controller(&state).await;
                        if recovered && completed.deferred_cleanup
                            && let Err(error) = state.start_deferred_cleanup(completed.session_id.clone())
                        {
                            tracing::warn!(session_id = %completed.session_id, error = format!("{error:#}"), "could not continue cleanup after interrupted close");
                            state.push_notice(
                                &completed.session_id,
                                format!("Could not continue container storage cleanup: {error:#}"),
                            );
                        }
                    }
                }
                accepted = listener.accept() => {
                    progress.phase(ServingPhase::Client);
                    let (stream, peer) = accepted.context("accept Mjolnir daemon client")?;
                    if !peer.ip().is_loopback() {
                        tracing::warn!(%peer, "rejected non-loopback daemon client");
                        continue;
                    }
                    let metadata = metadata.clone();
                    let state = state.clone();
                    let cancellation = cancellation.clone();
                    client_tasks.spawn(async move {
                        if let Err(error) = serve_client(stream, metadata, state, cancellation).await {
                            tracing::debug!(error = format!("{error:#}"), "daemon client disconnected");
                        }
                    });
                }
                completed = client_tasks.join_next(), if !client_tasks.is_empty() => {
                    if let Some(Err(error)) = completed {
                        tracing::warn!(%error, "daemon client task failed");
                    }
                }
                update = update_feed.next() => {
                    progress.phase(ServingPhase::SessionUpdate);
                    // The feed says why it ended. A shutdown also stops the
                    // continuation service and closes this channel, so a
                    // closed channel is not evidence of a failure.
                    let Some(update) = update? else { break };
                    if let Some((detail, observed_updated_at)) =
                        state.missing_target_record(&update.session_id, &update.view)
                    {
                        let state = state.clone();
                        let session_id = update.session_id.clone();
                        client_tasks.spawn(async move {
                            if let Err(error) = state.persist_missing_target(
                                &session_id, detail, observed_updated_at,
                            ).await {
                                tracing::warn!(%session_id, %error, "could not persist missing worker target");
                                state.push_notice(&session_id, format!("Could not record missing session target: {error:#}"));
                            }
                        });
                    }
                    if let Some(publisher) = phone_publisher.as_ref()
                        && let Err(error) = publisher.try_publish(
                            update.session_id.clone(),
                            update.view.clone(),
                        )
                    {
                        tracing::warn!(%error, "phone session view bridge stopped");
                        phone_publisher = None;
                    }
                    // Every session's view passes here whether or not anything is
                    // attached, which is exactly what an automatic review needs to
                    // see: the turn that just finished.
                    // The continuation completion gate has already notified review.
                    state.publish_session(update.session_id, update.view).await?;
                    // State is published before waking the delivery owner, so
                    // its readiness check observes the same session update.
                    crate::mailbox_outbox::notify_mailbox_outbox_changed();
                }
            }
        }
        Ok(())
    }
    .await;

    progress.phase(ServingPhase::Shutdown);
    let mut epilogue = EpilogueClock::start();
    epilogue_started.store(true, Ordering::Release);
    spawn_shutdown_watchdog();
    // Idle exit and fallible loop exits do not arrive through the termination
    // coordinator. Stop every daemon-owned task before closing the sole writer.
    cancellation.cancel();
    epilogue.record(
        &mut outcome,
        "join machine build cache applications",
        match cache_outcome {
            Some(result) => result,
            None => cache_task
                .await
                .context("machine cache service failed")
                .and_then(|result| result),
        },
    );
    epilogue.record(
        &mut outcome,
        "join mailbox outbox service",
        match mailbox_outcome {
            Some(result) => result,
            None => mailbox_task
                .await
                .context("mailbox outbox service failed")
                .and_then(|result| result),
        },
    );
    epilogue.record(
        &mut outcome,
        "join GitHub watch service",
        match github_watch_outcome {
            Some(result) => result,
            None => github_watch_task
                .await
                .context("GitHub watch service failed")
                .and_then(|result| result),
        },
    );
    // Recovery copies and worker upgrades do not hold up a handoff, so some
    // may still be running. Stop them first: the next daemon starts them again.
    // The coordinators share one gate, and dropping either cancels both.
    epilogue.record(
        &mut outcome,
        "shut down worker upgrade coordinator",
        worker_upgrades.shutdown().await,
    );
    epilogue.record(
        &mut outcome,
        "shut down recovery coordinator",
        recovery.shutdown().await,
    );
    epilogue.record(
        &mut outcome,
        "join continuation service",
        update_feed.join().await,
    );
    epilogue.record(
        &mut outcome,
        "join delegation coordinator",
        match delegation_outcome {
            Some(result) => result,
            None => delegation_task
                .await
                .map_err(anyhow::Error::from)
                .and_then(|result| result),
        },
    );
    drop(interrupted_close_tx);
    epilogue.record(
        &mut outcome,
        "remove daemon metadata",
        remove_daemon_metadata(&daemon_metadata_path),
    );
    epilogue.record(
        &mut outcome,
        "shut down turn review host",
        state
            .review_host()
            .shutdown()
            .await
            .map_err(anyhow::Error::msg),
    );
    epilogue.record(
        &mut outcome,
        "join CPU publication",
        cpu_publication.await.map_err(anyhow::Error::from),
    );
    record_daemon_cleanup(
        &mut outcome,
        "join project discovery",
        project_catalog.await.map_err(anyhow::Error::from),
    );
    epilogue.record(
        &mut outcome,
        "join controller target refresher",
        target_refresh.await.map_err(anyhow::Error::new),
    );
    epilogue.record(
        &mut outcome,
        "join container image refresher",
        image_refresh.await.map_err(anyhow::Error::new),
    );
    epilogue.record(
        &mut outcome,
        "join capacity service",
        capacity_service.await.map_err(anyhow::Error::new),
    );
    if let Some(phone_task) = phone_task {
        epilogue.record(
            &mut outcome,
            "join phone server",
            phone_task.await.map_err(anyhow::Error::new),
        );
    }
    if let Some(remote_request_bridge) = remote_request_bridge {
        epilogue.record(
            &mut outcome,
            "join phone session request bridge",
            remote_request_bridge.await.map_err(anyhow::Error::new),
        );
    }
    client_tasks.abort_all();
    while let Some(result) = client_tasks.join_next().await {
        if let Err(error) = result
            && !error.is_cancelled()
        {
            record_daemon_cleanup(
                &mut outcome,
                "join daemon client task",
                Err(anyhow::Error::new(error)),
            );
        }
    }
    epilogue.record(&mut outcome, "join daemon client tasks", Ok(()));
    epilogue.record(
        &mut outcome,
        "cancel daemon lifecycle operations",
        state.cancel_and_wait_lifecycles().await,
    );
    epilogue.record(
        &mut outcome,
        "drain startup prompts",
        state.cancel_and_join_startup_prompts().await,
    );
    if let Some(reconciliation) = reconciliation {
        epilogue.record(
            &mut outcome,
            "join interrupted lifecycle reconciliation",
            reconciliation.await.map_err(anyhow::Error::new),
        );
    }
    if let Some(tombstone_sweep) = tombstone_sweep {
        epilogue.record(
            &mut outcome,
            "join lost-session discard sweep",
            tombstone_sweep.await.map_err(anyhow::Error::new),
        );
    }
    epilogue.record(
        &mut outcome,
        "join checkpoint leftover sweep",
        checkpoint_sweep.await.map_err(anyhow::Error::new),
    );
    for interrupted_close_task in interrupted_close_tasks {
        epilogue.record(
            &mut outcome,
            "join interrupted close recovery",
            interrupted_close_task.await.map_err(anyhow::Error::new),
        );
    }
    epilogue.record(
        &mut outcome,
        "shut down controller daemon session manager",
        manager_shutdown.shutdown().await,
    );
    epilogue.finish();
    outcome
}

/// Log a startup phase that took long enough to matter, so a slow start after
/// an upgrade records where its time went.
fn log_startup_phase(phase: &'static str, started: Instant) {
    let took = started.elapsed();
    if took >= EpilogueClock::LOGGED_STEP {
        tracing::info!(
            phase,
            duration_ms = took.as_millis(),
            "daemon startup phase finished"
        );
    }
}

/// Sessions whose record is nothing but a tombstone: they ended without a
/// target and without a checkpoint, so the record offers no action but its own
/// removal. `DestroyedWithDataLoss` only ever comes from an older build.
pub(super) fn tombstone_session_ids(controller: &Controller) -> Vec<String> {
    controller
        .state
        .sessions
        .values()
        .filter(|session| {
            matches!(
                session.state,
                SessionState::Lost | SessionState::DestroyedWithDataLoss
            )
        })
        .map(|session| session.id.clone())
        .collect()
}

pub(super) fn remove_daemon_metadata(path: &Path) -> Result<()> {
    match fs::remove_file(path) {
        Ok(()) => Ok(()),
        Err(error) if error.kind() == std::io::ErrorKind::NotFound => Ok(()),
        Err(error) => Err(error).with_context(|| format!("remove {}", path.display())),
    }
}

/// Keep the event-loop failure as the primary result while still running and
/// reporting every cleanup step. If the loop ended normally, the first
/// cleanup failure becomes the daemon's result.
/// Times the shutdown epilogue, so a slow exit, such as the end of an upgrade
/// handoff, records which step took the time.
struct EpilogueClock {
    began: Instant,
    step: Instant,
}

impl EpilogueClock {
    /// Steps shorter than this are not logged.
    const LOGGED_STEP: Duration = Duration::from_millis(250);

    fn start() -> Self {
        let now = Instant::now();
        tracing::info!("daemon shutdown began");
        Self {
            began: now,
            step: now,
        }
    }

    /// Record `cleanup`, the result of the step that just finished. Steps run
    /// one after another, so the time since the previous one is this step's.
    fn record(&mut self, outcome: &mut Result<()>, operation: &'static str, cleanup: Result<()>) {
        let took = self.step.elapsed();
        if took >= Self::LOGGED_STEP {
            tracing::info!(
                step = operation,
                duration_ms = took.as_millis(),
                "daemon shutdown step finished"
            );
        }
        record_daemon_cleanup(outcome, operation, cleanup);
        self.step = Instant::now();
    }

    fn finish(self) {
        tracing::info!(
            duration_ms = self.began.elapsed().as_millis(),
            "daemon shutdown finished"
        );
    }
}

pub(super) fn record_daemon_cleanup(
    outcome: &mut Result<()>,
    operation: &'static str,
    cleanup: Result<()>,
) {
    let Err(error) = cleanup else {
        return;
    };
    let error = error.context(operation);
    if outcome.is_ok() {
        *outcome = Err(error);
    } else {
        tracing::warn!(error = format!("{error:#}"), "daemon cleanup step failed");
    }
}

/// Bounds the epilogue below.
///
/// The daemon leaves on its own long before this fires: a graceful exit
/// returns from `run_daemon_process`, the process exits 0, and this task dies
/// with the runtime. It exists so no unwinding step can hold the process open
/// past the deadline its clients wait on, whatever the cause of the shutdown.
pub(super) fn spawn_shutdown_watchdog() {
    tokio::spawn(async move {
        tokio::time::sleep(SHUTDOWN_FORCE_EXIT_TIMEOUT).await;
        tracing::error!(
            seconds = SHUTDOWN_FORCE_EXIT_TIMEOUT.as_secs(),
            "daemon shutdown did not finish in time; exiting"
        );
        // The metadata file points clients at a process that is about to stop
        // answering. Removing it is what the epilogue would have done.
        if let Err(error) = fs::remove_file(metadata_path())
            && error.kind() != std::io::ErrorKind::NotFound
        {
            tracing::warn!(%error, "could not remove daemon metadata before the forced exit");
        }
        // 128 + signal is reserved for exits that really were signalled.
        std::process::exit(1);
    });
}

pub(super) fn spawn_manager_target_refresher(
    targets: tokio::sync::watch::Sender<Vec<crate::session_manager::RelaySessionTarget>>,
    cancellation: CancellationToken,
    state: Arc<RuntimeState>,
) -> tokio::task::JoinHandle<()> {
    tokio::spawn(async move {
        let mut committed = state
            .committed
            .as_ref()
            .expect("daemon owns the writer")
            .clone();
        let mut revisions = state.revisions();
        let mut compatibility = tokio::time::interval(Duration::from_millis(500));
        compatibility.set_missed_tick_behavior(tokio::time::MissedTickBehavior::Delay);
        let mut installed = None;
        let mut credential_installed = None;
        let mut auth = crate::pollers::ProviderAuthCache::default();
        let mut auth_profiles = None;
        let mut staged_sources = None;
        loop {
            let inputs = {
                let owner = state.owner();
                if let Err(error) = owner.ensure_available() {
                    tracing::error!(%error, "daemon durable state is unavailable; shutting down");
                    cancellation.cancel();
                    return;
                }
                owner.pollable_worker_inputs()
            };
            let config = inputs.controller().config;
            if auth_profiles.as_ref() != Some(&config.profiles) {
                let mut cache = auth.clone();
                let profiles = config.profiles.clone();
                match tokio::task::spawn_blocking(move || {
                    cache.refresh(&config);
                    cache
                })
                .await
                {
                    Ok(cache) => {
                        auth = cache;
                        auth_profiles = Some(profiles);
                    }
                    Err(error) => {
                        tracing::error!(%error, "provider refresh task stopped");
                        cancellation.cancel();
                        return;
                    }
                }
            }
            if !installed
                .as_ref()
                .is_some_and(|current| Arc::ptr_eq(current, &inputs))
            {
                let preparation = inputs.clone();
                let refreshed =
                    match tokio::task::spawn_blocking(move || preparation.prepare()).await {
                        Ok(targets) => targets,
                        Err(error) => {
                            tracing::error!(%error, "worker target preparation failed");
                            cancellation.cancel();
                            return;
                        }
                    };
                let retained = refreshed
                    .iter()
                    .map(|target| target.session_id.clone())
                    .collect::<BTreeSet<_>>();
                let views_dropped = {
                    let mut owner = state.owner();
                    // A lifecycle may have claimed a target while its commands
                    // were being prepared. Only the owner can authorize install.
                    if !Arc::ptr_eq(&owner.pollable_worker_inputs(), &inputs) {
                        continue;
                    }
                    targets.send_if_modified(|current| {
                        if *current == refreshed {
                            false
                        } else {
                            *current = refreshed;
                            true
                        }
                    });
                    // Same lock as the install, so a view is published only
                    // while its actor is still wanted.
                    owner.install_relay_sessions(retained.clone())
                };
                if views_dropped {
                    state.publish_revision();
                }
                state.review_host().retain_sessions(retained);
                installed = Some(inputs.clone());
            }
            let closing = state.owner().close_requested.clone();
            if !staged_sources
                .as_ref()
                .is_some_and(|(current, _)| Arc::ptr_eq(current, &inputs))
            {
                let staging = inputs.clone();
                match tokio::task::spawn_blocking(move || staging.staged_credentials()).await {
                    Ok(staged) => staged_sources = Some((inputs.clone(), staged)),
                    Err(error) => {
                        tracing::error!(%error, "credential staging observation stopped");
                        cancellation.cancel();
                        return;
                    }
                }
            }
            let staged = &staged_sources
                .as_ref()
                .expect("credential sources observed")
                .1;
            let credential_key = (
                inputs.clone(),
                auth.schemes.clone(),
                closing.clone(),
                staged.clone(),
            );
            if !credential_installed.as_ref().is_some_and(
                |(current, schemes, closed, previous_staged)| {
                    Arc::ptr_eq(current, &inputs)
                        && *schemes == auth.schemes
                        && *closed == closing
                        && previous_staged == staged
                },
            ) {
                let preparation = inputs.clone();
                let schemes = auth.schemes.clone();
                let staged = staged.clone();
                let prepared = match tokio::task::spawn_blocking(move || {
                    preparation.prepare_credentials(&schemes, &staged)
                })
                .await
                {
                    Ok(targets) => targets,
                    Err(error) => {
                        tracing::error!(%error, "credential target preparation stopped");
                        cancellation.cancel();
                        return;
                    }
                };
                if !state.owner().install_credentials(
                    &inputs,
                    &closing,
                    prepared,
                    &state.credential_targets,
                ) {
                    continue;
                }
                credential_installed = Some(credential_key);
            }
            tokio::select! {
                _ = cancellation.cancelled() => return,
                changed = committed.changed() => {
                    if changed.is_err() {
                        tracing::error!("database publication feed stopped");
                        cancellation.cancel();
                        return;
                    }
                    state.publish_revision();
                }
                changed = revisions.changed() => {
                    if changed.is_err() { return; }
                }
                _ = compatibility.tick() => {
                    let _config_mutation = state.config_mutation.lock().await;
                    let mut cache = auth.clone();
                    let staging = inputs.clone();
                    let refreshed = tokio::task::spawn_blocking(move || {
                        crate::database::check_read_compatibility()?;
                        let config = Config::load()?;
                        cache.refresh(&config);
                        let staged = staging.staged_credentials();
                        Ok::<_, anyhow::Error>((config, cache, staged))
                    }).await;
                    match refreshed {
                        Ok(Ok((config, cache, staged))) => {
                            staged_sources = Some((inputs.clone(), staged));
                            auth = cache;
                            auth_profiles = Some(config.profiles.clone());
                            state.review_config.lock().unwrap_or_else(PoisonError::into_inner)
                                .clone_from(&config.review);
                            let changed = {
                                let mut owner = state.owner();
                                let changed = owner.controller().config != config;
                                owner.install_config(config);
                                changed
                            };
                            if changed { state.publish_revision(); }
                        }
                        Ok(Err(error)) => {
                            if error.chain().any(|cause| cause.downcast_ref::<StoreSchemaMismatch>().is_some()) {
                                tracing::error!(%error, "daemon store schema diverged underneath the daemon; shutting down");
                                cancellation.cancel();
                                return;
                            }
                            tracing::warn!(error = format!("{error:#}"), "could not refresh daemon configuration");
                        }
                        Err(error) => {
                            tracing::error!(%error, "daemon configuration reader failed");
                            cancellation.cancel();
                            return;
                        }
                    }
                }
            }
        }
    })
}

pub(super) async fn refresh_runtime_controller(state: &RuntimeState) {
    if let Err(error) = state.reload_controller().await {
        tracing::warn!(
            error = format!("{error:#}"),
            "could not refresh daemon controller state"
        );
    }
}