acme-proxy-server 0.6.0

//! The server runtime: how a configuration becomes routers, and how those are
//! served and rebuilt.
//!
//! Startup is split on the socket boundary, which is what lets a test drive the
//! whole path on an ephemeral port with its own shutdown future instead of a
//! process signal:
//!
//! - [`run`] binds `server.bind_address`, installs the `SIGHUP` handler, and
//!   hands the socket on. It is the whole of `acme-proxy serve`.
//! - [`serve_on`] validates the admin configuration and binds that socket too,
//!   when `[admin]` is enabled.
//! - [`serve_on_with`] does everything else — profile resolution, deduplicated
//!   signer backends, per-profile filters and validators, TLS, the job registry
//!   (every signer's handlers, notification delivery and the six table sweeps,
//!   built by `generation::job_registry_for`), the runner draining it, and
//!   `axum::serve` with connect info attached.
//!
//! That assembly is [`generation::build_generation`], and it is called again on
//! every reload rather than only at startup — so the two cannot drift, and a
//! subsystem added to one is added to the other by construction. What a reload
//! may change, and what it refuses by name, is [`crate::reload`]'s to say;
//! [`serve_on_with_reloads`] is where the two meet.
//!
//! - [`profile`] — how a generation builds each ACME endpoint.
//! - [`assembly`] — what a generation hands its profiles, and what outlives it.
//! - [`generation`] — one generation built, then published: the reload policy.
//! - [`supervisor`] — the task that serialises reloads.
//! - [`sockets`] — the three listeners' binds, plans and announcements.
//! - [`roles`] — which of `acme`, `admin` and `worker` this process runs.
//! - [`logging`] — the subscriber `serve` installs, and its reloadable filter.
//!
//! What it serves sits below it: the endpoint itself is
//! [`acme_proxy_protocol::profile::Profile`], and the routers each listener serves, with
//! their shared layers, are [`acme_proxy_protocol::router`].
//!
//! [`reload`] beside it is what a reload may change and what it refuses by
//! name. An internal crate of the `acme-proxy` binary, published in lockstep
//! with it and with no semver promise of its own.
use std::future::Future;
use std::net::SocketAddr;
use std::sync::Arc;

use tokio::net::TcpListener;
use tracing::{error, info, warn};

use acme_proxy_core::config::Config;
use acme_proxy_store::db::Database;

pub mod assembly;
pub mod generation;
pub mod logging;
pub mod profile;
pub mod reload;
pub mod roles;
pub mod sockets;
pub mod supervisor;
#[cfg(test)]
mod tests;

pub use assembly::{Assembly, GenerationParts};
pub use roles::{ProcessRole, RoleSet};

pub use sockets::check_metrics_config;

use generation::{Generation, announce_profile, build_generation};
use sockets::{
    Sockets, announce_admin_listener, announce_metrics_listener, bind_admin, bind_metrics,
    bound_address,
};
use supervisor::{Cells, supervise_reloads};

/// Binds the configured sockets and runs the server until a shutdown signal
/// arrives: the whole of `acme-proxy serve`.
///
/// Every failure is logged here — `server_socket_bind_failed` for the ACME
/// socket, `server_fatal_error` for anything after it — and returned for the
/// caller to turn into an exit status.
pub async fn run(
    roles: RoleSet,
    config: Arc<Config>,
    database: Arc<Database>,
) -> anyhow::Result<()> {
    // Only the roles this process runs get a socket. The ACME listener is the
    // one that used to be unconditional — there is no `server.enabled`, a CA
    // serving no ACME having been a process with nothing to do — and a role
    // split is exactly the case where that stops being true.
    let listener = match roles.has(ProcessRole::Acme) {
        false => None,
        true => Some(
            TcpListener::bind(&config.server.bind_address)
                .await
                .map_err(|error| {
                    error!(event = "server_socket_bind_failed",
                           outcome = "failure",
                           bind_address = %config.server.bind_address,
                           error = %error);
                    anyhow::anyhow!("cannot bind {}: {error}", config.server.bind_address)
                })?,
        ),
    };

    // Installed here, before anything slow: `SIGHUP`'s default disposition is
    // *terminate*, so until the handler exists a reload signal kills the
    // process. `serve_on_with_reloads` does profile assembly and the relay's
    // first upstream contact before it binds anything, which is exactly the
    // window an operator's `systemctl reload` could land in.
    let (reload_handle, reloads) = crate::reload::channel();
    let _hangups = AbortOnDrop(tokio::spawn(watch_for_hangup(reload_handle)));

    let fatal = |error: &anyhow::Error| {
        error!(event = "server_fatal_error", outcome = "failure", error = %error);
    };
    let admin_listener = match roles.has(ProcessRole::Admin) {
        false => None,
        true => bind_admin(&config).await.inspect_err(fatal)?,
    };
    let metrics_listener = bind_metrics(&config).await.inspect_err(fatal)?;

    serve_on_with_reloads(
        roles,
        config,
        database,
        Sockets {
            acme: listener,
            admin: admin_listener,
            metrics: metrics_listener,
        },
        shutdown_signal(),
        reloads,
    )
    .await
    .inspect_err(fatal)
}

/// Turns every `SIGHUP` into a reload request, for the life of the process.
///
/// Unlike the shutdown signal, this one does not consume its stream: an operator
/// reloads repeatedly, and a handler that fired once would leave the second
/// `SIGHUP` back at its default disposition — killing the server.
#[cfg(unix)]
async fn watch_for_hangup(handle: crate::reload::ReloadHandle) {
    let mut hangups = match tokio::signal::unix::signal(tokio::signal::unix::SignalKind::hangup()) {
        Ok(stream) => stream,
        Err(error) => {
            error!(event = "server_signal_handler_failed", outcome = "failure", signal = "SIGHUP", error = %error);
            return;
        }
    };
    while hangups.recv().await.is_some() {
        // `false` means the supervisor is gone — it panicked, or the server is
        // already shutting down — so this signal, and every later one, does
        // nothing. Silence there is an operator whose `systemctl reload`
        // reports success and changes nothing.
        if !handle.trigger() {
            error!(
                event = "server_reload_supervisor_gone",
                outcome = "failure",
                signal = "SIGHUP",
                "the reload supervisor is not accepting requests; restart the process to \
                 apply a configuration change"
            );
        }
    }
}

/// No `SIGHUP` off Unix, so there is nothing to watch for.
#[cfg(not(unix))]
async fn watch_for_hangup(_handle: crate::reload::ReloadHandle) {
    std::future::pending::<()>().await;
}

/// Assembles and serves the application over an already-bound socket.
///
/// Split from [`run`] on the socket boundary: a caller supplying its own
/// listener and its own `shutdown` future can drive the whole startup path —
/// profile assembly, TLS, backend resume, the nonce reaper, both `axum::serve`
/// arms — without owning a fixed port or a process signal.
///
/// Binds the web admin socket itself when `[admin]` is enabled. The signature
/// is unchanged, and `admin.enabled` is false by default, so every existing
/// caller is untouched; a test that wants to drive *both* listeners supplies
/// its own pair through [`serve_on_with`].
pub async fn serve_on(
    config: Arc<Config>,
    database: Arc<Database>,
    listener: TcpListener,
    shutdown: impl Future<Output = ()> + Send + 'static,
) -> anyhow::Result<()> {
    let admin_listener = bind_admin(&config).await?;
    let metrics_listener = bind_metrics(&config).await?;
    serve_on_with(
        config,
        database,
        listener,
        admin_listener,
        metrics_listener,
        shutdown,
    )
    .await
}

/// [`serve_on`] with all three sockets supplied.
///
/// The full version, split on the same boundary and for the same reason: a
/// caller handing in three ephemeral ports can drive the whole startup path —
/// including that one shutdown signal stops all of them — without owning a
/// fixed port or a process signal.
pub async fn serve_on_with(
    config: Arc<Config>,
    database: Arc<Database>,
    listener: TcpListener,
    admin_listener: Option<TcpListener>,
    metrics_listener: Option<TcpListener>,
    shutdown: impl Future<Output = ()> + Send + 'static,
) -> anyhow::Result<()> {
    serve_on_with_reloads(
        RoleSet::default(),
        config,
        database,
        Sockets {
            acme: Some(listener),
            admin: admin_listener,
            metrics: metrics_listener,
        },
        shutdown,
        crate::reload::Reloads::none(),
    )
    .await
}

/// [`serve_on_with`], serving configuration reloads as well as requests.
///
/// The variant [`run`] uses, so a `SIGHUP` rebuilds both routers, the job
/// registry, the notifier map and both TLS acceptors behind the sockets that are
/// already bound. Every other caller goes through [`serve_on_with`] and gets a
/// source that never fires, which costs one task that ends immediately.
///
/// See [`crate::reload`] for what a reload may change and what it refuses.
pub async fn serve_on_with_reloads(
    roles: RoleSet,
    config: Arc<Config>,
    database: Arc<Database>,
    sockets: Sockets,
    shutdown: impl Future<Output = ()> + Send + 'static,
    reloads: crate::reload::Reloads,
) -> anyhow::Result<()> {
    let Sockets {
        acme: listener,
        admin: admin_listener,
        metrics: metrics_listener,
    } = sockets;

    info!(
        event = "server_startup",
        outcome = "success",
        roles = %roles.labels().join(","),
        bind_address = %config.server.bind_address,
        base_url = %config.server.base_url,
        tls = config.server.tls.enabled,
        // A PostgreSQL DSN carries `user:password@`; this line is the one
        // place the whole value is printed at INFO on every start.
        database_database_url = %acme_proxy_core::logfields::redact_url(&config.database.url)
    );

    // The schema, before anything reads a row. **One owner**: the `worker` role
    // applies the migrations, and every other role checks and refuses by name.
    // Opening the database used to do this as a side effect, which made every
    // subcommand an upgrade step and let two processes starting together race
    // `MIGRATOR::run` with no lock between them.
    apply_or_require_schema(roles, &database).await?;

    // One `shutdown` future, several consumers: both listeners and the job
    // runner. Created here rather than beside `axum::serve` below so a signal
    // arriving *during* startup is not ignored — profile assembly and the
    // relay's first upstream contact both happen before anything binds. The
    // relay task is held under `AbortOnDrop` so an error path below does not
    // leak a task parked on a signal that will never arrive.
    let (shutdown_tx, shutdown_rx) = tokio::sync::watch::channel(false);
    let _shutdown_relay = AbortOnDrop(tokio::spawn(async move {
        shutdown.await;
        let _ = shutdown_tx.send(true);
    }));

    // The enqueue side of the durable queue. Built before the profiles because
    // a signer backend that defers issuance is handed one at construction, and
    // process-wide for the reason `[audit]` is: one table, one runner, and a
    // per-endpoint retry budget would make a job's pacing depend on which
    // profile happened to queue it.
    let job_queue = acme_proxy_jobs::jobs::JobQueue::new(database.clone(), &config.jobs);

    let resolved = config.resolve_profiles().inspect_err(|error| {
        error!(event = "profile_init_failed", outcome = "failure", error = %error);
    })?;
    // Everything that outlives a configuration generation — the signer backends
    // above all, which are carried rather than rebuilt. See
    // `crate::Assembly`.
    let (assembly, parts) = Assembly::new(
        roles,
        &resolved,
        database.clone(),
        job_queue.clone(),
        &config,
    )
    .inspect_err(|error| {
        error!(event = "profile_init_failed", outcome = "failure", error = %error);
    })?;
    let assembly = Arc::new(assembly);
    if roles.has(ProcessRole::Worker) {
        store_first_crls(&parts.signers).await;
    }

    let generation = build_generation(roles, &config, &resolved, &assembly, &parts, None)
        .inspect_err(|error| {
            error!(event = "profile_init_failed", outcome = "failure", error = %error);
        })?;

    // Every endpoint in the first generation came up, so every one is announced.
    // A reload announces only the endpoints it *added* — see
    // `supervise_reloads`, which shares this function for exactly that reason.
    for profile in &generation.profiles {
        announce_profile(profile).await;
    }

    let Generation {
        profiles: _,
        acme_app,
        admin_app,
        job_registry,
        tls,
        admin_tls,
        logins,
    } = generation;

    // The one task that drains the queue. Held under `AbortOnDrop` for the same
    // reason the reapers are — an un-cancelled loop holding an `Arc<Database>`
    // per `serve_on` call — and, unlike the relay tasks this replaced, it does
    // *not* need to outlive this function: the work is durable now, so a job cut
    // short is re-claimed from its own row rather than lost. It takes the same
    // shutdown signal as both listeners, so a stop is graceful rather than an
    // abort: it releases its leases on the way out, and a restart therefore
    // re-claims its own work immediately instead of waiting one out. The wait
    // for that stop is at the end of this function — `AbortOnDrop` is the
    // backstop for the paths that return before it, where an abort is the only
    // thing on offer.
    //
    // Neither the registry it drains nor the `[jobs]` section it paces itself
    // from is a value: both are cells a reload republishes, so a changed
    // retention, a rebuilt notify map and a retuned lease or concurrency all
    // reach the runner without restarting it. `jobs.max_attempts` is the third
    // piece and does not come through here — it belongs to the enqueue side, so
    // it is published onto `job_queue` itself.
    //
    // Only the `worker` role runs it. The cells are created either way, so the
    // supervisor keeps one shape and a reload republishes into them whether or
    // not anything is draining here.
    let (registry_tx, registry_rx) = tokio::sync::watch::channel(Arc::new(job_registry));
    let (jobs_tx, jobs_rx) = tokio::sync::watch::channel(Arc::new(config.jobs.clone()));
    let mut job_runner = roles.has(ProcessRole::Worker).then(|| {
        AbortOnDrop(acme_proxy_jobs::jobs::spawn_runner_watching(
            job_queue,
            registry_rx,
            jobs_rx,
            shutdown_rx.clone(),
        ))
    });
    if !roles.has(ProcessRole::Worker) {
        // Advisory rather than a refusal: an operator may legitimately start an
        // `acme` process before the worker, and the rows it queues are durable.
        // But a deployment that never runs one issues nothing — every relayed
        // order, notification, sweep and challenge validation waits for ever —
        // so this has to be visible.
        warn!(
            event = "server_role_no_worker",
            outcome = "advisory",
            roles = %roles.labels().join(","),
            "this process runs no worker, so nothing here drains the job queue: \
             challenge validation, notifications and the periodic sweeps all wait for a \
             process started with `--role worker`"
        );
    }

    // Only when this process actually holds the ACME socket. A worker-only
    // process announcing an address it never bound would send an operator
    // looking for a listener that is somebody else's.
    if roles.has(ProcessRole::Acme) {
        info!(
            event = "server_listening",
            outcome = "success",
            bind_address = %config.server.bind_address,
            protocol = if tls.is_some() { "https" } else { "http" }
        );
    }

    // One accept loop per role, each owning a socket a reload can replace and a
    // TLS mode it can switch — see `acme_proxy_net::listener`. `axum::serve` below is
    // handed one of these instead of a `TcpListener` and therefore outlives
    // every rebind, which is what removes the listener from the list of things
    // only a restart can change.
    let admin_bound = bound_address(admin_listener.as_ref(), &config.admin.bind_address);
    let metrics_bound = bound_address(metrics_listener.as_ref(), &config.metrics.bind_address);
    let (acme_socket, acme_handle) = acme_proxy_net::listener::spawn("acme", listener, tls);
    let (admin_socket, admin_handle) =
        acme_proxy_net::listener::spawn("admin", admin_listener, admin_tls);
    let (metrics_socket, metrics_handle) =
        acme_proxy_net::listener::spawn("metrics", metrics_listener, None);

    // Behind a swap cell rather than served directly, so a configuration reload
    // can replace the whole router without the socket moving. The cell is what
    // `axum::serve` holds; `acme_app` itself is only ever generation one.
    let (acme_router_tx, acme_router_rx) = crate::reload::router_channel(acme_app);
    let acme = serve_role(
        crate::reload::swappable(acme_router_rx),
        acme_socket,
        shutdown_rx.clone(),
    );

    // Opened whether or not the panel is on, unlike the app inside it: with
    // `admin.enabled` reloadable, a cell created only when the panel starts
    // would be the one thing a reload turning it on could not reach. An empty
    // `Router` answers `404` to everything, which is also what the panel being
    // switched off later publishes here.
    let (admin_router_tx, admin_router_rx) =
        crate::reload::router_channel(admin_app.unwrap_or_default());
    let admin = serve_role(
        crate::reload::swappable(admin_router_rx),
        admin_socket,
        shutdown_rx.clone(),
    );
    if roles.has(ProcessRole::Admin) && config.admin.enabled {
        announce_admin_listener(&config, &database, &admin_bound).await;
    }

    // The third listener. Served directly rather than through a
    // `reload::router_channel` like the other two, and the asymmetry is
    // deliberate: this router has one route whose only state is the registry,
    // and the registry is carried across generations rather than rebuilt (see
    // `Assembly`), so a new generation could put nothing new in it.
    if config.metrics.enabled {
        announce_metrics_listener(&metrics_bound);
    }
    let metrics = serve_role(
        acme_proxy_protocol::router::metrics_app(assembly.metrics.clone()),
        metrics_socket,
        shutdown_rx,
    );

    // The supervisor owns every cell sender from here on, which is what makes it
    // the only writer: a generation is published by one task or by nobody.
    // Aborted on drop, so an error return below does not leave it parked on a
    // channel nothing will ever send to.
    let _reload_supervisor = AbortOnDrop(tokio::spawn(supervise_reloads(
        roles,
        reloads,
        config.clone(),
        resolved,
        assembly,
        Cells {
            acme_router: acme_router_tx,
            admin_router: admin_router_tx,
            job_registry: registry_tx,
            jobs: jobs_tx,
            acme: acme_handle,
            admin: admin_handle,
            metrics: metrics_handle,
        },
        logins,
    )));

    // Nothing is drained here any more. A notification in flight at shutdown is
    // a `notify_deliver` row, not a spawned task: the runner releases its lease
    // on the way out and whoever starts next claims it. That is what replaced a
    // best-effort five-second drain which still lost anything slower than it.
    tokio::try_join!(acme, admin, metrics)?;

    // The listeners are done; the runner may not be. It takes the same shutdown
    // signal, but returning here would drop its `AbortOnDrop` and cut its stop
    // short — and in a worker-only process, where the three futures above have
    // nothing to finish, that happens immediately. Waiting for it is what makes
    // the release of its leases real: without it a job in flight waits out its
    // whole lease before another process may claim it, and `job_runner_stopped`
    // is never logged. Bounded, because a handler that ignores the signal must
    // not hold the process open: the runner's own budget plus a margin.
    if let Some(mut runner) = job_runner.take()
        && tokio::time::timeout(RUNNER_SHUTDOWN_BUDGET, runner.take())
            .await
            .is_err()
    {
        warn!(
            event = "job_runner_shutdown_timed_out",
            outcome = "failure",
            budget_ms = acme_proxy_core::logfields::millis(RUNNER_SHUTDOWN_BUDGET),
            "the job runner did not stop in time; its leases expire on their own"
        );
    }
    Ok(())
}

/// How long [`serve_on`] waits for the job runner to stop once both listeners
/// have. The runner's own drain budget is five seconds, so this is that plus
/// room for the writes that release its leases.
const RUNNER_SHUTDOWN_BUDGET: std::time::Duration = std::time::Duration::from_secs(10);

/// Stores each local CA's first CRL before this process serves anything.
///
/// Every role serves `GET /crl` from the stored row, and the read side never
/// signs, so a CA nothing has met yet has no CRL to serve. The daily
/// `CrlSweepJob` stores one on its first pass, but that pass runs on the
/// runner's first tick — after the listeners are up — so an all-in-one server
/// would answer `500` for the first moments of its life. Doing it here, once,
/// in the process that holds the keys, closes that for every topology with a
/// worker. A failure is logged where it happened
/// (`local_ca_crl_initialization_failed`) and the sweep's first pass tries
/// again, so it never stops the process.
/// Once per CA, not once per profile: two profiles over one `[signer]` section
/// share a backend, and a third naming the same CA by another path is still
/// the same issuer. Refreshing twice would sign the same CRL twice and export
/// it twice.
pub(crate) async fn store_first_crls(signers: &acme_proxy_signer::SignerSet) {
    let mut done: std::collections::HashSet<String> = std::collections::HashSet::new();
    for (_, backend) in signers.by_profile() {
        if let Some(refresher) = backend.crl_refresher()
            && done.insert(refresher.issuer().to_string())
        {
            let _ = refresher.refresh().await;
        }
    }
}

/// **The `worker` role owns the schema.** Every other role checks and stops by
/// name, which is what removes the startup race: on `SQLite` `sqlx` takes no
/// migration lock, so two processes that both ran `MIGRATOR::run` could
/// interleave. (`PostgreSQL`'s advisory lock would serialize them, but a
/// process that is not the worker still has no business changing the schema.) Naming `acme-proxy migrate` in the refusal also means a split
/// deployment fails at the process that started too early rather than later, as
/// a missing table in a request.
///
/// All-in-one is unaffected: a default `serve` runs `worker`, so a fresh
/// database is migrated exactly as it always was.
async fn apply_or_require_schema(roles: RoleSet, database: &Arc<Database>) -> anyhow::Result<()> {
    if roles.has(ProcessRole::Worker) {
        return database.migrate().await.map_err(|error| {
            error!(event = "db_migration_failed", outcome = "failure", error = %error);
            anyhow::anyhow!("cannot apply the database migrations: {error}")
        });
    }

    let pending = database.pending_migrations().await.map_err(|error| {
        error!(event = "server_schema_check_failed", outcome = "failure", error = %error);
        anyhow::anyhow!("cannot read the database schema version: {error}")
    })?;
    if pending.is_empty() {
        return Ok(());
    }

    error!(
        event = "server_schema_behind",
        outcome = "failure",
        pending = pending.len(),
        roles = %roles.labels().join(","),
    );
    anyhow::bail!(
        "the database is {} migration(s) behind and this process does not run the `worker` \
         role, which owns the schema: run `acme-proxy migrate` (or start the worker) first",
        pending.len()
    )
}

/// Serves `app` on one role's socket until the process shuts down.
///
/// One shape for all three roles, where there used to be a boxed future per
/// listener per TLS arm: [`acme_proxy_net::listener::RoleSocket`] is the same type
/// whether the role is speaking TLS, speaking cleartext or — a socket having
/// been closed by a reload — not serving at all, so the four cases collapse into
/// this one call. Its future lives for the process: a rebind replaces what is
/// underneath it, never the `axum::serve` above.
fn serve_role(
    app: axum::Router,
    socket: acme_proxy_net::listener::RoleSocket,
    shutdown: tokio::sync::watch::Receiver<bool>,
) -> impl Future<Output = std::io::Result<()>> + Send {
    axum::serve(
        socket,
        app.into_make_service_with_connect_info::<SocketAddr>(),
    )
    .with_graceful_shutdown(on_shutdown(shutdown))
    .into_future()
}

/// A future that completes when the shutdown relay fires.
async fn on_shutdown(mut receiver: tokio::sync::watch::Receiver<bool>) {
    // An error means the sender was dropped, which only happens when the relay
    // task itself is gone — treat it as "shut down" rather than parking
    // forever.
    let _ = receiver.wait_for(|ready| *ready).await;
}

async fn shutdown_signal() {
    let ctrl_c = async {
        tokio::signal::ctrl_c()
            .await
            .expect("failed to install Ctrl+C handler");
    };

    #[cfg(unix)]
    let terminate = async {
        tokio::signal::unix::signal(tokio::signal::unix::SignalKind::terminate())
            .expect("failed to install signal handler")
            .recv()
            .await;
    };

    #[cfg(not(unix))]
    let terminate = std::future::pending::<()>();

    tokio::select! {
        _ = ctrl_c => {},
        _ = terminate => {},
    }
}

/// Aborts a background task when it goes out of scope.
struct AbortOnDrop(tokio::task::JoinHandle<()>);

impl AbortOnDrop {
    /// Takes the handle back, so the caller can await the task's own stop
    /// instead of aborting it. What is left behind is an already-finished
    /// handle, whose abort on drop is a no-op.
    fn take(&mut self) -> tokio::task::JoinHandle<()> {
        std::mem::replace(&mut self.0, tokio::spawn(std::future::ready(())))
    }
}

impl Drop for AbortOnDrop {
    fn drop(&mut self) {
        self.0.abort();
    }
}