Skip to main content

subc_daemon/
bootstrap.rs

1use std::{
2    collections::BTreeMap,
3    env,
4    error::Error,
5    ffi::OsString,
6    fmt, fs, io,
7    net::{IpAddr, Ipv4Addr, Ipv6Addr, SocketAddr},
8    path::{Path, PathBuf},
9    process,
10    time::Duration,
11};
12
13use fs4::{FileExt, TryLockError};
14use subc_protocol::PROTOCOL_VERSION;
15pub use subc_transport::user_connection_token;
16use subc_transport::{
17    authenticate_client, connection_file, generate_daemon_id, generate_key, write_atomic,
18    AuthError, ConnectionFileError, ConnectionInfo, Endpoint, SCHEMA_VERSION,
19};
20use tokio::{
21    net::{TcpListener, TcpStream},
22    task::{JoinError, JoinHandle},
23    time::{sleep, timeout},
24};
25use tracing::{error, info, warn};
26
27use crate::{
28    daemon_config::{self, ConfiguredModule, DaemonConfigError},
29    server::{serve_listeners, ServerAuth, ServerError},
30    supervise::HealthConfig,
31    ConnectedClients, ControlHandler, DaemonSelfWatchdog, DaemonSelfWatchdogConfig,
32    ForwardingTable, Registry, RestartPolicy, Router, Supervisor, SupervisorHandle,
33    SupervisorProcessLiveness,
34};
35use std::sync::Arc;
36
37pub const DEFAULT_SUBC_PORT: u16 = 8757;
38pub const SUBC_PORT_ENV: &str = "SUBC_PORT";
39use subc_transport::CONNECTION_FILE_NAME;
40const DAEMON_VERSION: &str = env!("CARGO_PKG_VERSION");
41const CONNECT_TIMEOUT: Duration = Duration::from_secs(2);
42const PROBE_AUTH_DEADLINE: Duration = Duration::from_secs(2);
43const START_LOCK_RETRIES: usize = 40;
44const START_LOCK_RETRY_DELAY: Duration = Duration::from_millis(25);
45
46#[derive(Debug, Clone, Copy, PartialEq, Eq)]
47pub enum ConnectionFileSource {
48    XdgRuntimeDir,
49    TempDirFallback,
50    Explicit,
51}
52
53impl ConnectionFileSource {
54    fn reason(self) -> &'static str {
55        match self {
56            Self::XdgRuntimeDir => "XDG_RUNTIME_DIR set and non-empty",
57            Self::TempDirFallback => "XDG_RUNTIME_DIR unset or empty",
58            Self::Explicit => "configured path",
59        }
60    }
61}
62
63impl fmt::Display for ConnectionFileSource {
64    fn fmt(&self, formatter: &mut fmt::Formatter<'_>) -> fmt::Result {
65        formatter.write_str(match self {
66            Self::XdgRuntimeDir => "xdg_runtime_dir",
67            Self::TempDirFallback => "temp_dir_fallback",
68            Self::Explicit => "explicit",
69        })
70    }
71}
72
73/// Runtime bootstrap configuration. Production uses the default fixed port and
74/// optional daemon-config override; tests pass port 0 to let the OS assign a free
75/// loopback port and discover it from the connection file.
76#[derive(Debug, Clone, Default)]
77struct AdmissionFactsConfig {
78    carrier_module_id: Option<String>,
79    targets: Option<Vec<String>>,
80}
81
82/// Controls where module cgroups are prepared.
83///
84/// In-process daemons default to [`Self::Disabled`] so they never derive a
85/// production location from the host process. The shipped daemon explicitly uses
86/// [`Self::Current`]. The `ck-subc` binary accepts
87/// `SUBC_CGROUP_PLACEMENT=disabled` for isolated test processes; an unset
88/// variable retains `Current`, and every other value is rejected at startup.
89/// Tests that exercise placement can own a [`Self::Root`].
90#[derive(Debug, Clone, Default, PartialEq, Eq)]
91pub enum CgroupPlacementConfig {
92    #[default]
93    Disabled,
94    Current,
95    Root(PathBuf),
96}
97
98#[derive(Debug, Clone)]
99pub struct BootstrapConfig {
100    pub operator_queue_wait: Duration,
101    pub operator_timeout: Duration,
102    pub operator_stuck_grace: Duration,
103    #[cfg(feature = "test-support")]
104    pub operator_script: Option<PathBuf>,
105    #[cfg(feature = "test-support")]
106    pub operator_events: Option<PathBuf>,
107    pub connection_file_path: PathBuf,
108    pub port: u16,
109    pub daemon_ver: String,
110    configured_modules: Vec<ConfiguredModule>,
111    storage_config: Option<daemon_config::StorageConfig>,
112    admission_facts: AdmissionFactsConfig,
113    /// Module ids whose scopes may carry the attributes that grant authority.
114    scope_authority_owners: Vec<String>,
115    privileged_operator_prompt: daemon_config::PrivilegedOperatorPrompt,
116    activity_config: daemon_config::ActivityConfig,
117    daemon_config_path: Option<PathBuf>,
118    configured_port: Option<u16>,
119    /// Daemon-wide route.bind relay budget in milliseconds (the fallback for
120    /// any module without a per-module override). `None` = built-in default
121    /// (12s — see `control::DEFAULT_ROUTE_BIND_RELAY_TIMEOUT`).
122    route_bind_relay_default_ms: Option<u64>,
123    reserved_capabilities: BTreeMap<String, String>,
124    watchdog_config: DaemonSelfWatchdogConfig,
125    connection_file_source: ConnectionFileSource,
126    /// Where module cgroups are prepared. Disabled unless a caller explicitly
127    /// opts in, because `Current` derives a host location from `/proc/self/cgroup`.
128    cgroup_placement: CgroupPlacementConfig,
129    /// Directory the supervisor writes per-module stdout/stderr capture files
130    /// into. `None` disables capture; the shipped binary supplies its real run
131    /// directory explicitly.
132    capture_logs_dir: Option<PathBuf>,
133    /// File the supervisor appends every module's terminal exits to, so exit
134    /// history survives a daemon restart. `None` keeps terminal history in
135    /// memory only (each module's ring). Absent by default for the same reason
136    /// as `capture_logs_dir`: an in-process daemon booted by a test must not
137    /// append its exits to the operator's real `terminals.jsonl`, where they
138    /// would show up in `ck module terminals`. The shipped binary supplies
139    /// `<run dir>/terminals.jsonl` explicitly.
140    terminal_journal_path: Option<PathBuf>,
141    /// Where the machine id is read from, or minted into when absent. `None`
142    /// serves no machine id. Absent by default for the same reason as
143    /// `capture_logs_dir`: an in-process daemon booted by a test must never
144    /// derive the operator's real data home and mint into it. The shipped binary
145    /// supplies `<data home>/cortexkit/machine-id` explicitly.
146    machine_id_path: Option<PathBuf>,
147    /// The live-children record: every supervised process this daemon has
148    /// running, kept so the next daemon can end the ones a crash left behind.
149    /// At startup, before any module is spawned, the previous daemon's record
150    /// here is swept. `None` keeps no record and sweeps nothing, for the same
151    /// reason as `capture_logs_dir`: an in-process daemon booted by a test
152    /// must never signal processes listed in the operator's real run
153    /// directory. The shipped binary supplies `<run dir>/live-children.json`.
154    live_children_path: Option<PathBuf>,
155    privacy_trampoline: Option<PathBuf>,
156}
157
158impl BootstrapConfig {
159    pub fn new(connection_file_path: impl Into<PathBuf>, port: u16) -> Self {
160        Self {
161            connection_file_path: connection_file_path.into(),
162            operator_queue_wait: Duration::from_secs(150),
163            operator_timeout: Duration::from_secs(120),
164            operator_stuck_grace: Duration::from_secs(10),
165            #[cfg(feature = "test-support")]
166            operator_script: None,
167            #[cfg(feature = "test-support")]
168            operator_events: None,
169            port,
170            daemon_ver: DAEMON_VERSION.to_owned(),
171            configured_modules: Vec::new(),
172            storage_config: None,
173            admission_facts: AdmissionFactsConfig::default(),
174            scope_authority_owners: daemon_config::default_scope_authority_owners(),
175            privileged_operator_prompt: daemon_config::PrivilegedOperatorPrompt::default(),
176            activity_config: daemon_config::ActivityConfig::default(),
177            daemon_config_path: None,
178            configured_port: None,
179            route_bind_relay_default_ms: None,
180            reserved_capabilities: BTreeMap::new(),
181            watchdog_config: DaemonSelfWatchdogConfig::default(),
182            connection_file_source: ConnectionFileSource::Explicit,
183            cgroup_placement: CgroupPlacementConfig::default(),
184            capture_logs_dir: None,
185            terminal_journal_path: None,
186            machine_id_path: None,
187            live_children_path: None,
188            privacy_trampoline: None,
189        }
190    }
191
192    /// Keep the live-children record at `path`, and at startup end the
193    /// processes a previous daemon recorded there that are still running.
194    /// Embedding daemons and tests pass a path inside their own fixture tree.
195    pub fn with_live_children_record(mut self, path: impl Into<PathBuf>) -> Self {
196        self.live_children_path = Some(path.into());
197        self
198    }
199
200    /// macOS module launches require an explicit executable implementing the
201    /// subc-os trampoline entry point. Other systems ignore this setting.
202    pub fn with_privacy_trampoline(mut self, path: impl Into<PathBuf>) -> Self {
203        self.privacy_trampoline = Some(path.into());
204        self
205    }
206
207    /// Serve the machine id stored at `path`, minting it there at startup when
208    /// the file is absent. Embedding daemons and tests pass a path inside their
209    /// own fixture tree.
210    pub fn with_machine_id_path(mut self, path: impl Into<PathBuf>) -> Self {
211        self.machine_id_path = Some(path.into());
212        self
213    }
214
215    /// Selects module cgroup placement. The default is disabled.
216    pub fn with_cgroup_placement(mut self, placement: CgroupPlacementConfig) -> Self {
217        self.cgroup_placement = placement;
218        self
219    }
220
221    /// Redirects per-module stdout/stderr capture files out of the real run
222    /// directory. Tests that start an in-process daemon must call this with a
223    /// path inside their fixture tree.
224    pub fn with_capture_logs_dir(mut self, dir: impl Into<PathBuf>) -> Self {
225        self.capture_logs_dir = Some(dir.into());
226        self
227    }
228
229    /// Redirects the daemon-private journal, allowing embedded daemons and tests
230    /// to keep their observations out of the operator's live run directory.
231    pub fn with_terminal_journal_path(mut self, path: impl Into<PathBuf>) -> Self {
232        self.terminal_journal_path = Some(path.into());
233        self
234    }
235
236    pub fn from_env() -> Result<Self, BootstrapError> {
237        Self::from_env_with_daemon_config_path(daemon_config::default_config_path())
238    }
239
240    /// The SHIPPED BINARY's config, which is the only caller that should capture
241    /// child output into the operator's real run directory.
242    ///
243    /// Kept separate from `from_env` deliberately: see `capture_logs_dir` and
244    /// `terminal_journal_path` on this struct for why an absent value must mean
245    /// NO CAPTURE and NO JOURNAL rather than the operator's real run directory.
246    /// A run directory that cannot be resolved (a relative data home) refuses
247    /// startup instead of landing under the working directory.
248    pub fn from_env_for_daemon_binary() -> Result<Self, BootstrapError> {
249        let run_dir = daemon_config::daemon_run_dir().map_err(BootstrapError::RunDir)?;
250        let machine_id_path =
251            crate::machine_id::default_machine_id_path().map_err(BootstrapError::MachineId)?;
252        let config = Self::from_env()?
253            .with_capture_logs_dir(run_dir.join("logs"))
254            .with_terminal_journal_path(run_dir.join("terminals.jsonl"))
255            .with_live_children_record(crate::live_children::record_path(&run_dir))
256            .with_machine_id_path(machine_id_path);
257        #[cfg(feature = "test-support")]
258        let config = {
259            let mut config = config;
260            config.operator_script =
261                std::env::var_os("SUBC_TEST_OPERATOR_SCRIPT").map(PathBuf::from);
262            config.operator_events =
263                std::env::var_os("SUBC_TEST_OPERATOR_EVENTS").map(PathBuf::from);
264            for (name, value) in [
265                (
266                    "SUBC_TEST_OPERATOR_QUEUE_WAIT_MS",
267                    &mut config.operator_queue_wait,
268                ),
269                (
270                    "SUBC_TEST_OPERATOR_TIMEOUT_MS",
271                    &mut config.operator_timeout,
272                ),
273                (
274                    "SUBC_TEST_OPERATOR_STUCK_MS",
275                    &mut config.operator_stuck_grace,
276                ),
277            ] {
278                if let Ok(ms) = std::env::var(name) {
279                    if let Ok(ms) = ms.parse::<u64>() {
280                        *value = Duration::from_millis(ms);
281                    }
282                }
283            }
284            config
285        };
286        Ok(config)
287    }
288
289    pub fn from_env_with_daemon_config_path(
290        daemon_config_path: impl AsRef<Path>,
291    ) -> Result<Self, BootstrapError> {
292        let daemon_config_path = daemon_config_path.as_ref().to_path_buf();
293        let (daemon_config, activity_config) =
294            daemon_config::load_with_activity(&daemon_config_path)
295                .map_err(BootstrapError::DaemonConfig)?
296                .map(|(config, activity)| (Some(config), activity))
297                .unwrap_or_default();
298        let config_port = daemon_config.as_ref().and_then(|config| config.port);
299        let storage_config = daemon_config
300            .as_ref()
301            .and_then(|config| config.storage.clone());
302        let admission_facts_carrier_module_id = daemon_config
303            .as_ref()
304            .and_then(|config| config.admission_facts_carrier_module_id.clone());
305        let admission_facts_targets = daemon_config
306            .as_ref()
307            .and_then(|config| config.admission_facts_targets.clone());
308        let scope_authority_owners = daemon_config
309            .as_ref()
310            .map(|config| config.scope_authority_owners.clone())
311            .unwrap_or_else(daemon_config::default_scope_authority_owners);
312        let privileged_operator_prompt = daemon_config
313            .as_ref()
314            .map(|config| config.privileged_operator_prompt)
315            .unwrap_or_default();
316        let route_bind_relay_default_ms = daemon_config
317            .as_ref()
318            .and_then(|config| config.route_bind_relay_timeout_ms);
319        let reserved_capabilities = daemon_config
320            .as_ref()
321            .map(|config| config.reserved_capabilities.clone())
322            .unwrap_or_default();
323        let configured_modules = daemon_config
324            .map(|config| config.modules)
325            .unwrap_or_default();
326
327        let port = match env::var(SUBC_PORT_ENV) {
328            Ok(raw) if !raw.trim().is_empty() => {
329                let port = raw
330                    .parse::<u16>()
331                    .map_err(|source| BootstrapError::InvalidPort { raw, source })?;
332                if let Some(config_port) = config_port {
333                    info!(
334                        env = SUBC_PORT_ENV,
335                        env_port = port,
336                        config_port,
337                        "SUBC_PORT overrides daemon config port"
338                    );
339                }
340                port
341            }
342            Ok(_) | Err(_) => config_port.unwrap_or(DEFAULT_SUBC_PORT),
343        };
344
345        let (connection_file_path, connection_file_source) =
346            connection_file_path_with_source(non_empty_os_var("XDG_RUNTIME_DIR"));
347        Ok(Self::new(connection_file_path, port)
348            .with_configured_modules(configured_modules)
349            .with_storage_config(storage_config)
350            .with_admission_facts_config(admission_facts_carrier_module_id, admission_facts_targets)
351            .with_scope_authority_owners(scope_authority_owners)
352            .with_privileged_operator_prompt(privileged_operator_prompt)
353            .with_activity_config(activity_config)
354            .with_route_bind_relay_default_ms(route_bind_relay_default_ms)
355            .with_reserved_capabilities(reserved_capabilities)
356            .with_daemon_config_source(daemon_config_path, config_port)
357            .with_connection_file_source(connection_file_source))
358    }
359
360    pub fn with_daemon_config_path(
361        self,
362        daemon_config_path: impl AsRef<Path>,
363    ) -> Result<Self, BootstrapError> {
364        let daemon_config_path = daemon_config_path.as_ref().to_path_buf();
365        let (daemon_config, activity_config) =
366            daemon_config::load_with_activity(&daemon_config_path)
367                .map_err(BootstrapError::DaemonConfig)?
368                .map(|(config, activity)| (Some(config), activity))
369                .unwrap_or_default();
370        let configured_port = daemon_config.as_ref().and_then(|config| config.port);
371        let storage_config = daemon_config
372            .as_ref()
373            .and_then(|config| config.storage.clone());
374        let admission_facts_carrier_module_id = daemon_config
375            .as_ref()
376            .and_then(|config| config.admission_facts_carrier_module_id.clone());
377        let admission_facts_targets = daemon_config
378            .as_ref()
379            .and_then(|config| config.admission_facts_targets.clone());
380        let scope_authority_owners = daemon_config
381            .as_ref()
382            .map(|config| config.scope_authority_owners.clone())
383            .unwrap_or_else(daemon_config::default_scope_authority_owners);
384        let privileged_operator_prompt = daemon_config
385            .as_ref()
386            .map(|config| config.privileged_operator_prompt)
387            .unwrap_or_default();
388        let route_bind_relay_default_ms = daemon_config
389            .as_ref()
390            .and_then(|config| config.route_bind_relay_timeout_ms);
391        let reserved_capabilities = daemon_config
392            .as_ref()
393            .map(|config| config.reserved_capabilities.clone())
394            .unwrap_or_default();
395        let configured_modules = daemon_config
396            .map(|config| config.modules)
397            .unwrap_or_default();
398        Ok(self
399            .with_configured_modules(configured_modules)
400            .with_storage_config(storage_config)
401            .with_admission_facts_config(admission_facts_carrier_module_id, admission_facts_targets)
402            .with_scope_authority_owners(scope_authority_owners)
403            .with_privileged_operator_prompt(privileged_operator_prompt)
404            .with_activity_config(activity_config)
405            .with_route_bind_relay_default_ms(route_bind_relay_default_ms)
406            .with_reserved_capabilities(reserved_capabilities)
407            .with_daemon_config_source(daemon_config_path, configured_port))
408    }
409
410    pub fn with_configured_modules(
411        mut self,
412        modules: impl IntoIterator<Item = ConfiguredModule>,
413    ) -> Self {
414        self.configured_modules = modules.into_iter().collect();
415        self.configured_modules
416            .sort_by(|left, right| left.module_id.cmp(&right.module_id));
417        self
418    }
419
420    pub fn with_storage_config(
421        mut self,
422        storage_config: Option<daemon_config::StorageConfig>,
423    ) -> Self {
424        self.storage_config = storage_config;
425        self
426    }
427
428    pub fn with_admission_facts_config(
429        mut self,
430        carrier_module_id: Option<String>,
431        targets: Option<Vec<String>>,
432    ) -> Self {
433        self.admission_facts = AdmissionFactsConfig {
434            carrier_module_id,
435            targets,
436        };
437        self
438    }
439
440    /// Set the module ids whose scopes may carry `agent_id` and `delegates`.
441    /// Absent from a config, this is `daemon_config::default_scope_authority_owners`.
442    pub fn with_scope_authority_owners(mut self, owners: Vec<String>) -> Self {
443        self.scope_authority_owners = owners;
444        self
445    }
446
447    /// Choose the privileged-operation policy for the daemon's entire lifetime.
448    pub fn with_privileged_operator_prompt(
449        mut self,
450        mode: daemon_config::PrivilegedOperatorPrompt,
451    ) -> Self {
452        self.privileged_operator_prompt = mode;
453        self
454    }
455
456    pub fn with_activity_config(mut self, config: daemon_config::ActivityConfig) -> Self {
457        self.activity_config = config;
458        self
459    }
460
461    /// Set the daemon-wide route.bind relay default (the fallback for any
462    /// module without a per-module override). `None` preserves the built-in
463    /// default (12s). `serve_bound_daemon` reads this at startup and threads
464    /// it into the control handler's daemon-wide field.
465    pub fn with_route_bind_relay_default_ms(mut self, ms: Option<u64>) -> Self {
466        self.route_bind_relay_default_ms = ms;
467        self
468    }
469
470    pub fn with_reserved_capabilities(
471        mut self,
472        reserved_capabilities: BTreeMap<String, String>,
473    ) -> Self {
474        self.reserved_capabilities = reserved_capabilities;
475        self
476    }
477
478    fn with_daemon_config_source(
479        mut self,
480        daemon_config_path: PathBuf,
481        configured_port: Option<u16>,
482    ) -> Self {
483        self.daemon_config_path = Some(daemon_config_path);
484        self.configured_port = configured_port;
485        self
486    }
487
488    fn with_connection_file_source(mut self, source: ConnectionFileSource) -> Self {
489        self.connection_file_source = source;
490        self
491    }
492
493    pub fn with_watchdog_config(mut self, watchdog_config: DaemonSelfWatchdogConfig) -> Self {
494        self.watchdog_config = watchdog_config;
495        self
496    }
497}
498
499/// Result of singleton discovery.
500// Built once per daemon start and immediately matched, so the size gap between
501// the variants costs nothing worth a box on the public shape.
502#[allow(clippy::large_enum_variant)]
503#[derive(Debug)]
504pub enum Outcome {
505    /// A live daemon authenticated from the connection file; this invocation should exit 0.
506    AlreadyRunning,
507    /// This process won the singleton race, owns bound loopback listener(s), and
508    /// has published a fresh connection file.
509    Bound(BoundDaemon),
510}
511
512#[derive(Debug)]
513pub struct BoundDaemon {
514    pub listeners: Vec<TcpListener>,
515    pub connection_info: ConnectionInfo,
516    pub connection_file_path: PathBuf,
517    pub connection_file_source: ConnectionFileSource,
518    /// The machine id this daemon serves, established before any connection is
519    /// accepted. `None` when the config named no machine id path.
520    pub machine_id: Option<crate::machine_id::MachineId>,
521    /// Ownership of the run directory, held until the daemon exits so no other
522    /// daemon sweeps or writes this one's run state. `None` when the config
523    /// keeps no live-children record, and so has no run directory to own.
524    run_dir_lock: Option<crate::run_dir_lock::RunDirLock>,
525    /// The singleton start lock, kept when binding deferred publishing the
526    /// connection file. It stays held until the daemon has verified it is serving
527    /// and written the connection file, so a second daemon starting meanwhile
528    /// cannot claim the singleton too. `None` when the connection
529    /// file was already written at bind time.
530    publication_lock: Option<StartLock>,
531}
532
533/// Resolve subc's per-user TCP connection-file path.
534///
535/// `$XDG_RUNTIME_DIR/subc-connection.json` is preferred because the runtime
536/// directory is already per-user on Unix desktops. Without it, subc falls back
537/// to the system temp dir with a per-user token in the filename so different OS
538/// users do not collide on shared temp directories.
539pub fn connection_file_path() -> PathBuf {
540    connection_file_path_with_source(non_empty_os_var("XDG_RUNTIME_DIR")).0
541}
542
543fn connection_file_path_with_source(
544    runtime_dir: Option<OsString>,
545) -> (PathBuf, ConnectionFileSource) {
546    if let Some(runtime_dir) = runtime_dir.filter(|value| !value.is_empty()) {
547        return (
548            PathBuf::from(runtime_dir).join(CONNECTION_FILE_NAME),
549            ConnectionFileSource::XdgRuntimeDir,
550        );
551    }
552
553    (
554        env::temp_dir().join(format!("subc-{}.connection.json", user_connection_token())),
555        ConnectionFileSource::TempDirFallback,
556    )
557}
558
559/// Resolve, claim, and serve the per-user daemon singleton.
560///
561/// A second invocation is successful: if a live daemon authenticates from the
562/// existing connection file, this returns `Ok(())` after logging and the caller
563/// exits with status 0.
564pub async fn run() -> Result<(), BootstrapError> {
565    // `run` is a binary entry point, so it opts into the production cgroup and
566    // child-log locations. In-process callers keep both features disabled by default.
567    run_with_config(
568        BootstrapConfig::from_env_for_daemon_binary()?
569            .with_cgroup_placement(CgroupPlacementConfig::Current),
570    )
571    .await
572}
573
574/// Serve a daemon from an explicit config. This is the entry point the twelve
575/// sibling repos use to boot an in-process daemon in their integration tests.
576///
577/// # THIS INSTALLS NO TRACING SUBSCRIBER, SO THE DAEMON IS SILENT BY DEFAULT
578///
579/// The daemon's own diagnostics go through `tracing`, and `tracing` DISCARDS
580/// every event when no subscriber is installed. The shipped binary installs one
581/// in `main` (`init_tracing`); this function deliberately does not, because a
582/// library that installs a global subscriber fights with whatever the host
583/// process already set up.
584///
585/// The consequence for a test harness is not "less verbose": it is that the
586/// daemon has NOTHING TO SAY about any failure, and a missing instrument reads
587/// exactly like a clean one. PLEX found this on 2026-09-18 while trying to
588/// capture daemon logs beside an intermittent bind failure, and discovered the
589/// daemon had been silent in every conformance run that repo had ever done --
590/// so the one client-side error string was all the evidence that could exist,
591/// and they had spent a real investigation on a failure whose second source was
592/// never being recorded.
593///
594/// Install one in the harness before calling this, and assert it did something
595/// (they measured 236 daemon lines with the subscriber installed, 0 with the
596/// call commented out) -- otherwise the fix is itself unverified.
597pub async fn run_with_config(config: BootstrapConfig) -> Result<(), BootstrapError> {
598    let forwarding = Arc::new(ForwardingTable::default());
599    forwarding.operator_confirms().configure_bootstrap(&config);
600    let configured_modules = config.configured_modules.clone();
601    let storage_config = config.storage_config.clone();
602    let admission_facts = config.admission_facts.clone();
603    let scope_authority_owners = config.scope_authority_owners.clone();
604    let privileged_operator_prompt = config.privileged_operator_prompt;
605    let activity_config = config.activity_config;
606    let daemon_config_path = config.daemon_config_path.clone();
607    let configured_port = config.configured_port;
608    let route_bind_relay_default_ms = config.route_bind_relay_default_ms;
609    let reserved_capabilities = config.reserved_capabilities.clone();
610    let watchdog_config = config.watchdog_config.clone();
611    let cgroup_placement_config = config.cgroup_placement.clone();
612    let capture_logs_dir = config.capture_logs_dir.clone();
613    let terminal_journal_path = config.terminal_journal_path.clone();
614    let live_children_path = config.live_children_path.clone();
615    let privacy_trampoline = config.privacy_trampoline.clone();
616    match ensure_singleton_inner(config, false).await? {
617        Outcome::AlreadyRunning => {
618            info!("subc daemon already running");
619            Ok(())
620        }
621        Outcome::Bound(bound) => {
622            #[cfg(target_os = "linux")]
623            let cgroup_placement = prepare_cgroup_placement(&cgroup_placement_config);
624            #[cfg(not(target_os = "linux"))]
625            let _ = cgroup_placement_config;
626            serve_bound_daemon(
627                bound,
628                forwarding,
629                configured_modules,
630                storage_config,
631                admission_facts,
632                scope_authority_owners,
633                privileged_operator_prompt,
634                activity_config,
635                daemon_config_path,
636                configured_port,
637                route_bind_relay_default_ms,
638                reserved_capabilities,
639                watchdog_config,
640                capture_logs_dir,
641                terminal_journal_path,
642                live_children_path,
643                privacy_trampoline,
644                #[cfg(target_os = "linux")]
645                cgroup_placement,
646            )
647            .await
648        }
649    }
650}
651
652#[cfg(target_os = "linux")]
653fn prepare_cgroup_placement(config: &CgroupPlacementConfig) -> Option<subc_cgroup::Placement> {
654    let result = match config {
655        CgroupPlacementConfig::Disabled => return None,
656        CgroupPlacementConfig::Current => subc_cgroup::prepare_current(),
657        CgroupPlacementConfig::Root(root) => subc_cgroup::prepare_at(root),
658    };
659
660    match result {
661        Ok(Some(placement)) => Some(placement),
662        Ok(None) => {
663            warn!(
664                placement = ?config,
665                "module cgroup placement is disabled: configured cgroup root is not delegated"
666            );
667            None
668        }
669        Err(error) => {
670            warn!(
671                placement = ?config,
672                error = %error,
673                "module cgroup placement is disabled by an unexpected cgroup probe error"
674            );
675            None
676        }
677    }
678}
679
680/// Target soft limit for open file descriptors, applied to the daemon before any
681/// module is spawned so children inherit it. Multi-root modules (one process
682/// aggregating every project root's sqlite stores, index caches, watchers, and
683/// LSP pipes) trivially exceed the macOS default soft limit of 256; a launchd
684/// user agent does not pass login-shell ulimits through, so the raise must
685/// happen in-process.
686#[cfg(unix)]
687const NOFILE_TARGET: u64 = 65536;
688
689/// Raise RLIMIT_NOFILE to `NOFILE_TARGET` (clamped to the hard limit).
690/// Best-effort: failure is logged and never fatal, since the daemon can run
691/// under the inherited limit — modules with few roots just have less headroom.
692#[cfg(unix)]
693fn raise_nofile_limit() {
694    #[cfg(target_os = "macos")]
695    let ceiling = std::process::Command::new("/usr/sbin/sysctl")
696        .args(["-n", "kern.maxfilesperproc"])
697        .output()
698        .ok()
699        .filter(|output| output.status.success())
700        .and_then(|output| String::from_utf8(output.stdout).ok())
701        .and_then(|value| value.trim().parse::<u64>().ok())
702        .filter(|value| *value > 0);
703    #[cfg(not(target_os = "macos"))]
704    let ceiling = None;
705    raise_nofile_limit_with_ceiling(ceiling);
706}
707
708#[cfg(unix)]
709fn raise_nofile_limit_with_ceiling(kernel_ceiling: Option<u64>) {
710    match rlimit::Resource::NOFILE.get() {
711        Ok((soft, hard)) => {
712            // On XNU an infinite hard limit does not remove the separate
713            // per-process kernel ceiling. Exceeding it makes setrlimit fail
714            // without raising the inherited (often 256) soft limit at all.
715            let target = NOFILE_TARGET
716                .min(hard)
717                .min(kernel_ceiling.unwrap_or(u64::MAX));
718            if soft >= target {
719                return;
720            }
721            match rlimit::Resource::NOFILE.set(target, hard) {
722                Ok(()) => info!(
723                    previous_soft = soft,
724                    new_soft = target,
725                    hard,
726                    "raised open-file soft limit for daemon and module children"
727                ),
728                Err(err) => warn!(
729                    soft,
730                    hard,
731                    error = %err,
732                    "could not raise open-file soft limit; multi-root modules may exhaust descriptors"
733                ),
734            }
735        }
736        Err(err) => warn!(error = %err, "could not read open-file limit"),
737    }
738}
739
740/// CRT stdio-stream target on Windows (the `_setmaxstdio` maximum). Win32
741/// HANDLEs — what Rust `File`, tokio sockets, and SQLite's Win32 VFS actually
742/// consume — have a per-process quota in the millions and need no raise; the
743/// C-runtime stream table (default 512) is the only low ceiling, and it is
744/// per-process rather than inherited, so supervised modules linking the CRT
745/// must raise their own. Raising it here covers the daemon itself.
746#[cfg(windows)]
747fn raise_nofile_limit() {
748    const MAXSTDIO_TARGET: u32 = 8192;
749    let current = rlimit::getmaxstdio();
750    if current >= MAXSTDIO_TARGET {
751        return;
752    }
753    match rlimit::setmaxstdio(MAXSTDIO_TARGET) {
754        Ok(new_max) => info!(
755            previous = current,
756            new_max, "raised CRT stdio-stream limit for daemon"
757        ),
758        Err(err) => warn!(
759            current,
760            error = %err,
761            "could not raise CRT stdio-stream limit"
762        ),
763    }
764}
765
766#[cfg(not(any(unix, windows)))]
767fn raise_nofile_limit() {}
768
769pub async fn run_with_daemon_config_path(
770    config: BootstrapConfig,
771    daemon_config_path: impl AsRef<Path>,
772) -> Result<(), BootstrapError> {
773    run_with_config(config.with_daemon_config_path(daemon_config_path)?).await
774}
775
776#[allow(clippy::too_many_arguments)]
777async fn serve_bound_daemon(
778    bound: BoundDaemon,
779    forwarding: Arc<ForwardingTable>,
780    configured_modules: Vec<ConfiguredModule>,
781    storage_config: Option<daemon_config::StorageConfig>,
782    admission_facts: AdmissionFactsConfig,
783    scope_authority_owners: Vec<String>,
784    privileged_operator_prompt: daemon_config::PrivilegedOperatorPrompt,
785    activity_config: daemon_config::ActivityConfig,
786    daemon_config_path: Option<PathBuf>,
787    configured_port: Option<u16>,
788    route_bind_relay_default_ms: Option<u64>,
789    reserved_capabilities: BTreeMap<String, String>,
790    watchdog_config: DaemonSelfWatchdogConfig,
791    capture_logs_dir: Option<PathBuf>,
792    terminal_journal_path: Option<PathBuf>,
793    live_children_path: Option<PathBuf>,
794    privacy_trampoline: Option<PathBuf>,
795    #[cfg(target_os = "linux")] cgroup_placement: Option<subc_cgroup::Placement>,
796) -> Result<(), BootstrapError> {
797    #[cfg(unix)]
798    let mut terminate = tokio::signal::unix::signal(tokio::signal::unix::SignalKind::terminate())
799        .map_err(BootstrapError::Signal)?;
800    // Windows has no SIGTERM. Ctrl-C is a different event, not an equivalent
801    // service-stop contract, so this shutdown handler is intentionally Unix-only.
802    raise_nofile_limit();
803    #[cfg(windows)]
804    if let Err(error) = subc_jobobject::ensure_root_job() {
805        // Keep the control plane available. The root-job API caches setup
806        // failures, so later module spawns refuse instead of running unprotected.
807        warn!(%error, "module launches disabled: Windows root job setup failed");
808    }
809
810    info!(
811        connection_file = %bound.connection_file_path.display(),
812        connection_file_source = %bound.connection_file_source,
813        connection_file_source_reason = bound.connection_file_source.reason(),
814        endpoints = ?bound.connection_info.endpoints,
815        configured_modules = configured_modules.len(),
816        machine_id = bound.machine_id.as_ref().map(|id| id.as_str()).unwrap_or("none"),
817        "subc daemon starting"
818    );
819
820    // Before anything can spawn a module (the configured modules below, or a
821    // client's start request once the listeners are served): a previous daemon
822    // that died without its shutdown stop may have left children running, and
823    // a fresh copy beside one would fight it for its port and stores. The
824    // record is not a running daemon's because this daemon holds the lock on
825    // the run directory the record lives in, and a running daemon holds that
826    // lock for its whole life. (Claiming the singleton is not enough: it is
827    // keyed on the connection file, which can live in a different runtime
828    // directory from the run directory.) The lock stays held until this
829    // function returns, which is when the daemon exits.
830    let run_dir_lock = bound.run_dir_lock;
831    let publication_lock = bound.publication_lock;
832    if let Some(owner) = &run_dir_lock {
833        crate::live_children::sweep_orphans(
834            owner,
835            &crate::live_children::AdoptedPids::none(),
836            crate::live_children::SweepBounds::default(),
837        )
838        .await;
839    }
840
841    let registry = Arc::new(Registry::default());
842    let process_liveness = Arc::new(SupervisorProcessLiveness::new());
843    let supervisor_handle = SupervisorHandle::new();
844    let connected_clients = ConnectedClients::new();
845    let daemon_incarnation = format!(
846        "{:032x}",
847        u128::from_be_bytes(bound.connection_info.daemon_id)
848    );
849    let supervisor = Supervisor::new(Arc::clone(&registry), RestartPolicy::default())
850        .with_process_liveness(process_liveness.clone())
851        .with_forwarding(Arc::clone(&forwarding))
852        .with_handle(supervisor_handle.clone())
853        .with_connection_file_path(bound.connection_file_path.clone())
854        .with_daemon_incarnation(daemon_incarnation.clone());
855    let supervisor = match privacy_trampoline {
856        Some(path) => supervisor.with_privacy_trampoline(path),
857        None => supervisor,
858    };
859    // ABSENT MEANS NO CAPTURE AND NO JOURNAL, NOT "THE REAL RUN DIRECTORY", and
860    // the difference is a production-corruption hazard rather than a preference.
861    // Both fields below follow the same rule: `None` for `capture_logs_dir`
862    // means supervised output is not captured, and `None` for
863    // `terminal_journal_path` means terminal history lives only in each
864    // module's in-memory ring (`supervisor.terminals` still answers, with the
865    // journal counters at zero).
866    //
867    // The terminal journal used to fall back to
868    // `daemon_run_dir().join("terminals.jsonl")`, so an in-process test daemon
869    // appended its fixture exits to the operator's journal, where they then
870    // appeared in `ck module terminals`.
871    //
872    // The capture line used to be `unwrap_or_else(|| daemon_run_dir().join("logs"))`,
873    // so ANY caller that did not set the field captured supervised children into
874    // the operator's live `~/.local/share/cortexkit/run/logs/`. That is twelve
875    // sibling repos whose integration tests boot an in-process daemon through
876    // `run_with_config` -- none of which asked for it, and none of which can see
877    // it from their side.
878    //
879    // Harmless while fixture module ids are fixture-shaped: this host carries 20
880    // zero-byte files from subc's own tests (good-aft, missing-aft,
881    // preview-consumer...). THE HAZARD IS A COLLISION. A fixture named "broca"
882    // or "aft" appends to a PRODUCTION capture file that operators read
883    // forensically and that placement gates count lines in -- with no residue to
884    // notice, because the file legitimately exists and legitimately grows.
885    //
886    // Found by BROCA (2026-09-19) from the other side: their rigs spawn the
887    // SHIPPED ck-subc and set XDG_CONFIG_HOME + XDG_RUNTIME_DIR but not
888    // XDG_DATA_HOME, so every local rig run supervised a module named "broca"
889    // and captured it into production's broca.stderr.log -- the same file I
890    // count seal lines in before and after placing their binaries.
891    //
892    // The supervisor already treats `None` as no-capture and no-journal, so
893    // this only removes invented defaults. The binary keeps capturing and
894    // journaling via `BootstrapConfig::from_env_for_daemon_binary`.
895
896    // Keep both ledgers in the explicitly selected run directory. An embedded
897    // daemon without a journal path must not write to the operator's live ledger.
898    let activity_run_dir = terminal_journal_path
899        .as_ref()
900        .and_then(|path| path.parent())
901        .map(Path::to_path_buf);
902    let supervisor = match terminal_journal_path {
903        Some(path) => supervisor.with_terminal_journal(path, daemon_incarnation.clone()),
904        None => supervisor,
905    };
906    let supervisor = match capture_logs_dir {
907        Some(dir) => supervisor.with_capture_logs_dir(dir),
908        None => supervisor,
909    };
910    let supervisor = match live_children_path {
911        Some(path) => supervisor.with_live_children_record(path),
912        None => supervisor,
913    };
914    #[cfg(target_os = "linux")]
915    let supervisor = supervisor.with_cgroup_placement(cgroup_placement);
916    // Collect per-module route.bind relay overrides BEFORE handing the
917    // `configured_modules` vector to the supervisor (which only needs each
918    // module's `drain_timeout_ms`). Each entry was filled in by parse-time
919    // resolution (per-module > daemon-wide > absent), so modules with no
920    // override are absent from this map and the daemon-wide default applies.
921    let route_bind_relay_timeouts = configured_modules
922        .iter()
923        .filter_map(|module| {
924            module
925                .route_bind_relay_timeout_ms
926                .map(|ms| (module.module_id.clone(), Duration::from_millis(ms)))
927        })
928        .collect::<std::collections::BTreeMap<_, _>>();
929    let control_start_clock = crate::clock::StartClock::capture();
930    let mut control = ControlHandler::with_forwarding(Arc::clone(&registry), forwarding)
931        .with_process_liveness(process_liveness)
932        .with_supervisor(supervisor_handle)
933        .with_connected_clients(connected_clients.clone())
934        .with_storage_config(storage_config)
935        .with_machine_id(bound.machine_id.clone())
936        .with_admission_facts_config(admission_facts.carrier_module_id, admission_facts.targets)
937        .with_scope_authority_owners(scope_authority_owners)
938        .with_privileged_operator_prompt(privileged_operator_prompt)
939        .with_route_bind_relay_timeouts(route_bind_relay_timeouts)
940        .with_daemon_provenance(
941            bound.connection_info.pid,
942            control_start_clock.started_at_ms(),
943            std::env::current_exe().ok(),
944            normalized_build_provenance(env!("SUBC_BUILD_GIT_SHA")),
945            normalized_build_provenance(env!("SUBC_BUILD_LOCK_DIGEST")),
946        )
947        .with_daemon_start_clock(control_start_clock)
948        .with_capability_config(
949            configured_modules
950                .iter()
951                .map(|module| (module.module_id.clone(), module.enabled)),
952            reserved_capabilities,
953        );
954    if let Some(run_dir) = activity_run_dir {
955        control =
956            control.with_activity_ledger_config(&run_dir, daemon_incarnation, activity_config);
957    }
958    if let Some(ms) = route_bind_relay_default_ms {
959        // A daemon-wide config value overrides the built-in default; a
960        // `None` here leaves the ControlHandler's 12s default in place.
961        control = control.with_route_bind_relay_timeout(Duration::from_millis(ms));
962    }
963    if let Some(config_path) = daemon_config_path {
964        control = control.with_supervisor_rescan(supervisor.clone(), config_path, configured_port);
965    }
966    let control = Arc::new(control);
967    let router = Arc::new(Router::with_control_handler(Arc::clone(&control)));
968    let auth = ServerAuth::new(
969        bound.connection_info.key.clone(),
970        bound.connection_info.daemon_id,
971        bound.connection_info.daemon_ver.clone(),
972    )
973    .with_connected_clients(connected_clients);
974
975    // Discovery must expose every accepted configured module, including disabled
976    // ones. Registration creates status handles but neither starts monitors
977    // nor launches processes: children need the published file for `--subc`.
978    let mut registered_modules = Vec::with_capacity(configured_modules.len());
979    for configured in configured_modules {
980        let enabled = configured.enabled;
981        let health = configured.health.clone();
982        let module_id = configured.module_id.clone();
983        let registered = match supervisor.register_configured_with_health(
984            configured.module_spec(),
985            enabled,
986            health.clone(),
987            configured.drain_timeout_ms,
988            configured.restart,
989        ) {
990            Ok(registered) => registered,
991            Err(err) => {
992                error!(module_id = %module_id, error = %err, "failed to supervise configured module; continuing daemon startup");
993                continue;
994            }
995        };
996        registered_modules.push(registered);
997        // A relaxed threshold can silently outlive the temporary allowance
998        // that required it, so report the extra tolerance on every boot.
999        let default_threshold = HealthConfig::default().failure_threshold;
1000        if enabled && health.failure_threshold > default_threshold {
1001            warn!(
1002                module_id = %module_id,
1003                failure_threshold = health.failure_threshold,
1004                default_threshold,
1005                tolerance_secs = health.cadence.as_secs() * u64::from(health.failure_threshold),
1006                "health failure threshold is relaxed above the default; a wedged module stays unflagged for longer"
1007            );
1008        }
1009        info!(module_id = %module_id, enabled, "configured module registered");
1010    }
1011    control.refresh_capability_requirements();
1012
1013    let mut serve_task =
1014        AbortOnDrop::new(tokio::spawn(serve_listeners(bound.listeners, router, auth)));
1015    verify_serving(&bound.connection_info).await?;
1016    // A published endpoint promises that authentication can complete now.
1017    // Orphan cleanup and control-plane construction can take longer than an
1018    // existing-daemon probe's deadline, so finish both before discovery.
1019    // The write fsyncs, so it runs on tokio's blocking pool (see
1020    // `publish_connection_file`). The singleton lock travels with it and comes
1021    // back only once the file is in place.
1022    let (published, publication_lock) = publish_connection_file(
1023        &bound.connection_file_path,
1024        &bound.connection_info,
1025        publication_lock,
1026    )
1027    .await;
1028    published?;
1029    #[cfg(test)]
1030    startup_publication_hook::run(&bound.connection_file_path).await;
1031    drop(publication_lock);
1032    let _clock_step_task = AbortOnDrop::new(crate::watchdog::spawn_clock_step_monitor());
1033    // Off the startup path: it runs `systemctl`, and only ever logs.
1034    #[cfg(target_os = "linux")]
1035    let _kill_mode_check = AbortOnDrop::new(tokio::spawn(
1036        crate::systemd_kill_mode::warn_if_kill_mode_defeats_ordered_shutdown(),
1037    ));
1038    let _watchdog_task = AbortOnDrop::new(
1039        DaemonSelfWatchdog::new(
1040            bound.connection_info.clone(),
1041            bound.connection_file_path.clone(),
1042        )
1043        .with_config(watchdog_config)
1044        .spawn(),
1045    );
1046
1047    for registered in registered_modules {
1048        match supervisor.start_registered_configured(registered, None) {
1049            Ok(module) => {
1050                info!(module_id = %module.module_id(), "configured module supervised");
1051            }
1052            Err(err) => {
1053                error!(error = %err, "failed to start configured module; continuing daemon startup");
1054            }
1055        }
1056    }
1057
1058    control.refresh_capability_requirements();
1059    Arc::clone(&control).spawn_capability_deadline_loop();
1060    Arc::clone(&control).spawn_scope_expiry_loop();
1061
1062    #[cfg(unix)]
1063    {
1064        tokio::select! {
1065            result = serve_task.join() => {
1066                return result.map_err(BootstrapError::ServeJoin)?.map_err(BootstrapError::Serve);
1067            }
1068            _ = terminate.recv() => {}
1069        }
1070        // First, before the notice, the drain, or any connection close: from
1071        // here on a module exit is recorded as `daemon_shutdown` and never
1072        // respawned. Also before allowing a second signal to cut the bounded
1073        // wait short, so the journal marker is always written.
1074        supervisor.begin_daemon_shutdown();
1075        // Stop the self-watchdog before closing the listener. Its next tick would
1076        // connect to that listener, fail, and log an ERROR indistinguishable from
1077        // a wedged daemon, once for every interval a planned stop lasts.
1078        drop(_watchdog_task);
1079        // Dropping the listener stops new accepts, not established connections:
1080        // their detached tasks must remain live throughout notice and drain.
1081        drop(serve_task);
1082        let escalated = tokio::select! {
1083            biased;
1084            _ = terminate.recv() => {
1085                info!("second SIGTERM: abandoning daemon shutdown wait");
1086                true
1087            }
1088            result = supervisor.drain_for_daemon_shutdown() => {
1089                if let Err(error) = result {
1090                    warn!(%error, "daemon shutdown drain failed; exiting anyway");
1091                }
1092                false
1093            }
1094        };
1095        // Supervised modules lead their own process groups, so a service
1096        // manager's kill of this process's group does not reach them. (A
1097        // systemd unit with KillMode=control-group kills by cgroup instead and
1098        // does reach them; see `systemd_kill_mode`.) The daemon ends them
1099        // itself: EOF (or SIGTERM for a protocol none child) first, then
1100        // signals at each child's own drain deadline.
1101        supervisor
1102            .end_children_for_daemon_shutdown(escalated, async {
1103                terminate.recv().await;
1104            })
1105            .await;
1106        info!("daemon exiting: SIGTERM shutdown complete");
1107        Ok(())
1108    }
1109    #[cfg(not(unix))]
1110    serve_task
1111        .join()
1112        .await
1113        .map_err(BootstrapError::ServeJoin)?
1114        .map_err(BootstrapError::Serve)
1115}
1116
1117fn normalized_build_provenance(value: &str) -> Option<String> {
1118    match value.trim() {
1119        "" | "unavailable" => None,
1120        value => Some(value.to_string()),
1121    }
1122}
1123
1124/// Find an existing daemon or atomically bind loopback TCP for this daemon.
1125///
1126/// The algorithm is intentionally connect-first: an endpoint from the connection
1127/// file is treated as live only after the TCP+key server-proof authenticates for
1128/// that file's key and daemon_id. Stale or foreign connection files are reclaimed
1129/// only while holding the per-user start lock; the TCP port is never the
1130/// singleton primitive.
1131pub async fn ensure_singleton(
1132    connection_file_path: impl AsRef<Path>,
1133    port: u16,
1134) -> Result<Outcome, BootstrapError> {
1135    ensure_singleton_with_config(BootstrapConfig::new(connection_file_path.as_ref(), port)).await
1136}
1137
1138pub async fn ensure_singleton_with_config(
1139    config: BootstrapConfig,
1140) -> Result<Outcome, BootstrapError> {
1141    ensure_singleton_inner(config, true).await
1142}
1143
1144async fn ensure_singleton_inner(
1145    config: BootstrapConfig,
1146    publish: bool,
1147) -> Result<Outcome, BootstrapError> {
1148    let path = config.connection_file_path;
1149
1150    if matches!(probe_existing(&path).await?, Probe::Live) {
1151        return Ok(Outcome::AlreadyRunning);
1152    }
1153
1154    let lock = StartLock::acquire(&path).await?;
1155
1156    // Re-probe after acquiring the start lock so a peer that won the race between
1157    // our first failed probe and the lock acquisition is observed instead of
1158    // overwritten.
1159    if matches!(probe_existing(&path).await?, Probe::Live) {
1160        return Ok(Outcome::AlreadyRunning);
1161    }
1162
1163    // The start lock and the probe above only rule out a daemon using this
1164    // connection file. A daemon started with another runtime directory but
1165    // the same data home shares the run directory while passing both, so the
1166    // run directory needs its own owner: refuse at once if a live daemon holds
1167    // it, before anything below reads or writes run or data-home state.
1168    let run_dir_lock = config
1169        .live_children_path
1170        .as_deref()
1171        .map(crate::run_dir_lock::RunDirLock::acquire)
1172        .transpose()?;
1173
1174    remove_stale_connection_file_if_present(&path)?;
1175
1176    // Established under the start lock, and under the run-directory lock when
1177    // there is one, and before binding: a corrupt file stops boot before
1178    // anything is published, and two daemons racing to start cannot both mint.
1179    //
1180    // Minting a new machine id fsyncs its file, so loading or minting runs on
1181    // tokio's blocking pool. Both locks travel with that work, so neither is
1182    // released while a mint is in flight.
1183    let (machine_id, lock, run_dir_lock) = match config.machine_id_path {
1184        None => (None, lock, run_dir_lock),
1185        Some(machine_id_path) => {
1186            let task_path = machine_id_path.clone();
1187            match run_blocking_holding((lock, run_dir_lock), move || {
1188                crate::machine_id::load_or_mint(&task_path)
1189            })
1190            .await
1191            {
1192                Ok((Ok(id), (lock, run_dir_lock))) => (Some(id), lock, run_dir_lock),
1193                Ok((Err(error), _locks)) => return Err(BootstrapError::MachineId(error)),
1194                Err(join_error) => {
1195                    return Err(BootstrapError::MachineId(
1196                        crate::machine_id::MachineIdFileError::Write {
1197                            path: machine_id_path,
1198                            source: io::Error::other(describe_blocking_task_failure(
1199                                "loading or minting the machine id",
1200                                &join_error,
1201                            )),
1202                        },
1203                    ))
1204                }
1205            }
1206        }
1207    };
1208
1209    let (listeners, endpoints) = bind_loopback(config.port).await?;
1210    let connection_info = ConnectionInfo {
1211        schema: SCHEMA_VERSION,
1212        wire_version: Some(PROTOCOL_VERSION),
1213        endpoints,
1214        key: generate_key().map_err(BootstrapError::GenerateConnectionFile)?,
1215        daemon_id: generate_daemon_id().map_err(BootstrapError::GenerateConnectionFile)?,
1216        pid: process::id(),
1217        daemon_ver: config.daemon_ver,
1218    };
1219
1220    // The low-level binding API historically publishes for callers that serve
1221    // these listeners themselves. The daemon entry point defers publication
1222    // until its own authentication server is running.
1223    let publication_lock = if publish {
1224        let (published, lock) = publish_connection_file(&path, &connection_info, lock).await;
1225        if let Err(error) = published {
1226            drop(listeners);
1227            return Err(error);
1228        }
1229        drop(lock);
1230        None
1231    } else {
1232        Some(lock)
1233    };
1234
1235    Ok(Outcome::Bound(BoundDaemon {
1236        listeners,
1237        connection_info,
1238        connection_file_path: path,
1239        connection_file_source: config.connection_file_source,
1240        machine_id,
1241        run_dir_lock,
1242        publication_lock,
1243    }))
1244}
1245
1246#[derive(Debug, Clone, Copy, PartialEq, Eq)]
1247enum Probe {
1248    Live,
1249    StaleOrAbsent,
1250}
1251
1252async fn probe_existing(path: &Path) -> Result<Probe, BootstrapError> {
1253    let info = match connection_file::read(path) {
1254        Ok(info) => info,
1255        Err(source) if is_absent_or_stale_connection_file(&source) => {
1256            return Ok(Probe::StaleOrAbsent)
1257        }
1258        Err(source) => {
1259            return Err(BootstrapError::ConnectionFileRead {
1260                path: path.to_path_buf(),
1261                source,
1262            })
1263        }
1264    };
1265
1266    for endpoint in &info.endpoints {
1267        if matches!(probe_endpoint(&info, endpoint).await, Probe::Live) {
1268            return Ok(Probe::Live);
1269        }
1270    }
1271
1272    Ok(Probe::StaleOrAbsent)
1273}
1274
1275async fn probe_endpoint(info: &ConnectionInfo, endpoint: &Endpoint) -> Probe {
1276    let Ok(ip) = endpoint.host.parse::<IpAddr>() else {
1277        return Probe::StaleOrAbsent;
1278    };
1279    if !ip.is_loopback() {
1280        return Probe::StaleOrAbsent;
1281    }
1282    let addr = SocketAddr::new(ip, endpoint.port);
1283
1284    let mut stream = match timeout(CONNECT_TIMEOUT, TcpStream::connect(addr)).await {
1285        Ok(Ok(stream)) => stream,
1286        Ok(Err(_)) | Err(_) => return Probe::StaleOrAbsent,
1287    };
1288
1289    match authenticate_client(&mut stream, info, PROBE_AUTH_DEADLINE).await {
1290        Ok(()) => Probe::Live,
1291        Err(AuthError::DaemonIdMismatch)
1292        | Err(AuthError::InvalidServerProof)
1293        | Err(AuthError::UnexpectedEof { .. })
1294        | Err(AuthError::Timeout { .. })
1295        | Err(AuthError::JsonEncode { .. })
1296        | Err(AuthError::JsonDecode { .. })
1297        | Err(AuthError::Io { .. })
1298        | Err(AuthError::MessageTooLarge { .. })
1299        | Err(AuthError::KeyTooShort { .. })
1300        | Err(AuthError::Random(_))
1301        | Err(AuthError::InvalidClientAuth) => Probe::StaleOrAbsent,
1302    }
1303}
1304
1305async fn verify_serving(info: &ConnectionInfo) -> Result<(), BootstrapError> {
1306    // Scheduling a server task is not a readiness barrier. Authenticate once
1307    // against the mandatory IPv4 listener before exposing discovery to peers.
1308    // This uses the normal probe deadline, with no retries or widened budget.
1309    if let Some(endpoint) = info.endpoints.first() {
1310        if matches!(probe_endpoint(info, endpoint).await, Probe::Live) {
1311            return Ok(());
1312        }
1313    }
1314    Err(BootstrapError::StartupNotServing)
1315}
1316
1317/// Run blocking filesystem work (an fsync, a rename) on tokio's blocking-thread
1318/// pool rather than on the async worker thread that awaits it. On a slow or
1319/// busy disk an fsync can take seconds, and a runtime has few worker threads
1320/// (an in-process test daemon may share two with everything else), so doing it
1321/// inline stalls unrelated tasks for that long.
1322///
1323/// `guards` (lock handles) move into the blocking task and are handed back
1324/// with the work's result. If the awaiting future is dropped first, the work
1325/// still runs to completion on its thread and the guards are released only
1326/// after it ends, so a lock is never given up while the work it protects is
1327/// still writing.
1328///
1329/// `Err` means the task panicked, or was cancelled before it started because
1330/// the runtime is shutting down; the guards have then been released already.
1331async fn run_blocking_holding<G, T>(
1332    guards: G,
1333    work: impl FnOnce() -> T + Send + 'static,
1334) -> Result<(T, G), JoinError>
1335where
1336    G: Send + 'static,
1337    T: Send + 'static,
1338{
1339    tokio::task::spawn_blocking(move || {
1340        let output = work();
1341        (output, guards)
1342    })
1343    .await
1344}
1345
1346/// Explain, for a startup error, why blocking work produced no result.
1347fn describe_blocking_task_failure(work: &str, join_error: &JoinError) -> String {
1348    if join_error.is_panic() {
1349        format!("the background thread {work} panicked before finishing: {join_error}")
1350    } else {
1351        format!(
1352            "the background thread {work} was cancelled before finishing, because the async runtime is shutting down: {join_error}"
1353        )
1354    }
1355}
1356
1357/// Publish the connection file: owner-only permissions, written to a temporary
1358/// file, fsynced and renamed over `path` (all done by `write_atomic`). The
1359/// write runs on tokio's blocking pool through [`run_blocking_holding`], so the
1360/// fsync never holds an async worker thread.
1361///
1362/// `guard` is the singleton start lock (or whatever the caller must keep held
1363/// across publication); it is returned with the result. It comes back `None`
1364/// only when the write task panicked or was cancelled, in which case the
1365/// result is an error and the file must be treated as not published.
1366async fn publish_connection_file<G: Send + 'static>(
1367    path: &Path,
1368    info: &ConnectionInfo,
1369    guard: G,
1370) -> (Result<(), BootstrapError>, Option<G>) {
1371    let task_path = path.to_path_buf();
1372    let task_info = info.clone();
1373    let outcome =
1374        run_blocking_holding(guard, move || write_connection_file(&task_path, &task_info)).await;
1375    match outcome {
1376        Ok((Ok(()), guard)) => (Ok(()), Some(guard)),
1377        Ok((Err(source), guard)) => (
1378            Err(BootstrapError::ConnectionFileWrite {
1379                path: path.to_path_buf(),
1380                source,
1381            }),
1382            Some(guard),
1383        ),
1384        Err(join_error) => (
1385            Err(BootstrapError::ConnectionFileWrite {
1386                path: path.to_path_buf(),
1387                source: ConnectionFileError::Io {
1388                    op: "write",
1389                    path: path.to_path_buf(),
1390                    source: io::Error::other(describe_blocking_task_failure(
1391                        "writing the connection file",
1392                        &join_error,
1393                    )),
1394                },
1395            }),
1396            None,
1397        ),
1398    }
1399}
1400
1401/// The synchronous connection-file write, including its fsync. Call it only
1402/// through [`publish_connection_file`], which keeps it off the async workers.
1403fn write_connection_file(path: &Path, info: &ConnectionInfo) -> Result<(), ConnectionFileError> {
1404    #[cfg(test)]
1405    connection_file_write_hook::run(path);
1406    write_atomic(path, info)
1407}
1408
1409/// Lets a test run code inside the connection-file write, on whichever thread
1410/// performs it, for example to block there and check what else can still run.
1411#[cfg(test)]
1412mod connection_file_write_hook {
1413    use std::{
1414        path::{Path, PathBuf},
1415        sync::{Mutex, PoisonError},
1416    };
1417
1418    type Hook = Box<dyn FnOnce() + Send>;
1419
1420    /// Keyed by connection-file path, so a hook fires only for the test that
1421    /// installed it while other tests publish their own files in parallel.
1422    static HOOKS: Mutex<Vec<(PathBuf, Hook)>> = Mutex::new(Vec::new());
1423
1424    /// Run `hook` once, at the start of the next write of `path`.
1425    pub(super) fn install(path: &Path, hook: impl FnOnce() + Send + 'static) {
1426        HOOKS
1427            .lock()
1428            .unwrap_or_else(PoisonError::into_inner)
1429            .push((path.to_path_buf(), Box::new(hook)));
1430    }
1431
1432    pub(super) fn run(path: &Path) {
1433        let hook = {
1434            let mut hooks = HOOKS.lock().unwrap_or_else(PoisonError::into_inner);
1435            hooks
1436                .iter()
1437                .position(|(hooked, _)| hooked == path)
1438                .map(|index| hooks.swap_remove(index).1)
1439        };
1440        if let Some(hook) = hook {
1441            hook();
1442        }
1443    }
1444}
1445
1446/// A path-scoped barrier after publication and before any configured activation.
1447/// The async wait leaves the listeners free to answer discovery-time requests.
1448#[cfg(test)]
1449mod startup_publication_hook {
1450    use std::{
1451        path::{Path, PathBuf},
1452        sync::{Mutex, PoisonError},
1453    };
1454    use tokio::sync::oneshot;
1455
1456    type Barrier = (oneshot::Sender<()>, oneshot::Receiver<()>);
1457    static BARRIERS: Mutex<Vec<(PathBuf, Barrier)>> = Mutex::new(Vec::new());
1458
1459    pub(super) fn install(path: &Path) -> (oneshot::Receiver<()>, oneshot::Sender<()>) {
1460        let (entered_tx, entered_rx) = oneshot::channel();
1461        let (release_tx, release_rx) = oneshot::channel();
1462        BARRIERS
1463            .lock()
1464            .unwrap_or_else(PoisonError::into_inner)
1465            .push((path.to_path_buf(), (entered_tx, release_rx)));
1466        (entered_rx, release_tx)
1467    }
1468
1469    pub(super) async fn run(path: &Path) {
1470        let barrier = {
1471            let mut barriers = BARRIERS.lock().unwrap_or_else(PoisonError::into_inner);
1472            barriers
1473                .iter()
1474                .position(|(hooked, _)| hooked == path)
1475                .map(|index| barriers.swap_remove(index).1)
1476        };
1477        if let Some((entered, release)) = barrier {
1478            let _ = entered.send(());
1479            let _ = release.await;
1480        }
1481    }
1482}
1483
1484fn is_absent_or_stale_connection_file(err: &ConnectionFileError) -> bool {
1485    match err {
1486        ConnectionFileError::Io { source, .. } if source.kind() == io::ErrorKind::NotFound => true,
1487        ConnectionFileError::JsonRead { .. }
1488        | ConnectionFileError::UnsupportedSchema { .. }
1489        | ConnectionFileError::Invalid { .. }
1490        | ConnectionFileError::KeyTooShort { .. }
1491        // A live daemon always publishes the file owner-only (0600), so a file
1492        // with insecure permissions is never a daemon we should defer to: treat it
1493        // as stale and take over (which republishes a correct 0600 file).
1494        | ConnectionFileError::InsecurePermissions { .. } => true,
1495        ConnectionFileError::MissingParent { .. }
1496        | ConnectionFileError::MissingFileName { .. }
1497        // A writable ancestor is an operator misconfiguration, never evidence
1498        // about whether a daemon is live. Reclaiming the file would republish key
1499        // material into the same directory the refusal is about.
1500        | ConnectionFileError::InsecureParentDirectory { .. }
1501        | ConnectionFileError::Io { .. }
1502        | ConnectionFileError::JsonWrite { .. }
1503        | ConnectionFileError::Random(_)
1504        // A wire mismatch may identify a newer live daemon, so never reclaim its
1505        // connection file merely because this binary cannot speak its envelope.
1506        | ConnectionFileError::WireVersionMismatch { .. } => false,
1507    }
1508}
1509
1510fn remove_stale_connection_file_if_present(path: &Path) -> Result<(), BootstrapError> {
1511    match fs::remove_file(path) {
1512        Ok(()) => Ok(()),
1513        Err(err) if err.kind() == io::ErrorKind::NotFound => Ok(()),
1514        Err(source) => Err(BootstrapError::RemoveStale {
1515            path: path.to_path_buf(),
1516            source,
1517        }),
1518    }
1519}
1520
1521async fn bind_loopback(port: u16) -> Result<(Vec<TcpListener>, Vec<Endpoint>), BootstrapError> {
1522    bind_loopback_with(port, |port| TcpListener::bind((Ipv6Addr::LOCALHOST, port))).await
1523}
1524
1525async fn bind_loopback_with<F, Fut>(
1526    port: u16,
1527    mut bind_v6: F,
1528) -> Result<(Vec<TcpListener>, Vec<Endpoint>), BootstrapError>
1529where
1530    F: FnMut(u16) -> Fut,
1531    Fut: std::future::Future<Output = io::Result<TcpListener>>,
1532{
1533    let v4_host = Ipv4Addr::LOCALHOST;
1534    let v4 = TcpListener::bind((v4_host, port))
1535        .await
1536        .map_err(|source| BootstrapError::Bind {
1537            host: v4_host.to_string(),
1538            port,
1539            source,
1540        })?;
1541    let actual_port = v4
1542        .local_addr()
1543        .map_err(|source| BootstrapError::LocalAddr {
1544            host: v4_host.to_string(),
1545            source,
1546        })?
1547        .port();
1548
1549    let mut listeners = vec![v4];
1550    let mut endpoints = vec![Endpoint {
1551        host: v4_host.to_string(),
1552        port: actual_port,
1553    }];
1554
1555    let v6_host = Ipv6Addr::LOCALHOST;
1556    match bind_v6(actual_port).await {
1557        Ok(v6) => {
1558            listeners.push(v6);
1559            endpoints.push(Endpoint {
1560                host: v6_host.to_string(),
1561                port: actual_port,
1562            });
1563        }
1564        Err(err)
1565            if ipv6_loopback_unavailable(&err)
1566                || (port == 0 && err.kind() == io::ErrorKind::AddrInUse) =>
1567        {
1568            warn!(
1569                port = actual_port,
1570                error = %err,
1571                "IPv6 loopback unavailable; serving only IPv4 loopback"
1572            );
1573        }
1574        Err(source) => {
1575            drop(listeners);
1576            return Err(BootstrapError::Bind {
1577                host: v6_host.to_string(),
1578                port: actual_port,
1579                source,
1580            });
1581        }
1582    }
1583
1584    Ok((listeners, endpoints))
1585}
1586
1587fn ipv6_loopback_unavailable(err: &io::Error) -> bool {
1588    matches!(
1589        err.kind(),
1590        io::ErrorKind::AddrNotAvailable | io::ErrorKind::Unsupported
1591    ) || matches!(err.raw_os_error(), Some(47) | Some(49) | Some(97))
1592}
1593
1594struct AbortOnDrop<T> {
1595    handle: JoinHandle<T>,
1596}
1597
1598impl<T> AbortOnDrop<T> {
1599    fn new(handle: JoinHandle<T>) -> Self {
1600        Self { handle }
1601    }
1602
1603    async fn join(&mut self) -> Result<T, JoinError> {
1604        (&mut self.handle).await
1605    }
1606}
1607
1608impl<T> Drop for AbortOnDrop<T> {
1609    fn drop(&mut self) {
1610        if !self.handle.is_finished() {
1611            self.handle.abort();
1612        }
1613    }
1614}
1615
1616#[derive(Debug)]
1617struct StartLock {
1618    // Keep the locked file handle alive for the duration of bootstrap.
1619    _file: fs::File,
1620}
1621
1622impl Drop for StartLock {
1623    fn drop(&mut self) {
1624        // Closing alone leaves the lock held if a concurrent fork inherited
1625        // the open file description and has not exec'd yet. Unlock explicitly
1626        // so a restart can reclaim the singleton as soon as bootstrap ends.
1627        if let Err(error) = FileExt::unlock(&self._file) {
1628            tracing::warn!(%error, "failed to release the singleton start lock");
1629        }
1630    }
1631}
1632
1633impl StartLock {
1634    async fn acquire(connection_file_path: &Path) -> Result<Self, BootstrapError> {
1635        let path = start_lock_path(connection_file_path);
1636        for _ in 0..START_LOCK_RETRIES {
1637            let file = match open_owner_only_lock(&path) {
1638                Ok(file) => file,
1639                Err(source) => return Err(BootstrapError::StartLockCreate { path, source }),
1640            };
1641            match FileExt::try_lock(&file) {
1642                Ok(()) => return Ok(Self { _file: file }),
1643                Err(TryLockError::WouldBlock) => sleep(START_LOCK_RETRY_DELAY).await,
1644                Err(TryLockError::Error(source)) => {
1645                    return Err(BootstrapError::StartLockCreate { path, source });
1646                }
1647            }
1648        }
1649
1650        Err(BootstrapError::StartLockBusy {
1651            path,
1652            attempts: START_LOCK_RETRIES,
1653        })
1654    }
1655}
1656
1657pub(crate) fn open_owner_only_lock(path: &Path) -> io::Result<fs::File> {
1658    let mut options = fs::OpenOptions::new();
1659    options.read(true).write(true).create(true);
1660    #[cfg(unix)]
1661    {
1662        use std::os::unix::fs::OpenOptionsExt;
1663        options.mode(0o600);
1664    }
1665    options.open(path)
1666}
1667
1668fn start_lock_path(connection_file_path: &Path) -> PathBuf {
1669    let file_name = connection_file_path
1670        .file_name()
1671        .map(|name| name.to_string_lossy())
1672        .unwrap_or_else(|| CONNECTION_FILE_NAME.into());
1673    let lock_name = format!("{file_name}.start-lock");
1674    connection_file_path
1675        .parent()
1676        .filter(|parent| !parent.as_os_str().is_empty())
1677        .unwrap_or_else(|| Path::new("."))
1678        .join(lock_name)
1679}
1680
1681fn non_empty_os_var(key: &str) -> Option<OsString> {
1682    let value = env::var_os(key)?;
1683    if value.is_empty() {
1684        None
1685    } else {
1686        Some(value)
1687    }
1688}
1689
1690/// Bootstrap-layer errors are deliberately typed so startup never panics for
1691/// ordinary daemon-discovery races or stale filesystem state.
1692#[derive(Debug)]
1693pub enum BootstrapError {
1694    /// The bound server could not authenticate its own readiness probe.
1695    StartupNotServing,
1696    #[cfg(unix)]
1697    Signal(io::Error),
1698    InvalidPort {
1699        raw: String,
1700        source: std::num::ParseIntError,
1701    },
1702    ConnectionFileRead {
1703        path: PathBuf,
1704        source: ConnectionFileError,
1705    },
1706    ConnectionFileWrite {
1707        path: PathBuf,
1708        source: ConnectionFileError,
1709    },
1710    GenerateConnectionFile(ConnectionFileError),
1711    StartLockCreate {
1712        path: PathBuf,
1713        source: io::Error,
1714    },
1715    StartLockBusy {
1716        path: PathBuf,
1717        attempts: usize,
1718    },
1719    /// The run-directory lock file could not be created, opened or locked.
1720    RunDirLockCreate {
1721        path: PathBuf,
1722        source: io::Error,
1723    },
1724    /// Another live process, almost certainly a daemon started with a
1725    /// different runtime directory over the same data home, owns the run
1726    /// directory. The daemon does not start. `holder_pid` is read from the
1727    /// lock file when possible and is informational only.
1728    RunDirBusy {
1729        path: PathBuf,
1730        holder_pid: Option<u32>,
1731    },
1732    RemoveStale {
1733        path: PathBuf,
1734        source: io::Error,
1735    },
1736    Bind {
1737        host: String,
1738        port: u16,
1739        source: io::Error,
1740    },
1741    LocalAddr {
1742        host: String,
1743        source: io::Error,
1744    },
1745    DaemonConfig(DaemonConfigError),
1746    /// The machine id could not be established: its file is corrupt, unreadable
1747    /// or unwritable, or the data home is relative. The daemon does not start.
1748    MachineId(crate::machine_id::MachineIdFileError),
1749    /// The daemon run directory could not be resolved because the data home is
1750    /// relative. The daemon does not start.
1751    RunDir(daemon_config::DaemonRunDirError),
1752    Serve(ServerError),
1753    ServeJoin(tokio::task::JoinError),
1754}
1755
1756impl fmt::Display for BootstrapError {
1757    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
1758        match self {
1759            Self::StartupNotServing => write!(f, "bound daemon did not authenticate its readiness probe; connection file was not published"),
1760            #[cfg(unix)]
1761            Self::Signal(error) => write!(f, "failed to register SIGTERM handler: {error}"),
1762            Self::InvalidPort { raw, source } => {
1763                write!(f, "invalid {SUBC_PORT_ENV} value '{raw}': {source}")
1764            }
1765            Self::ConnectionFileRead { path, source } => write!(
1766                f,
1767                "failed to read connection file {}: {source}",
1768                path.display()
1769            ),
1770            Self::ConnectionFileWrite { path, source } => write!(
1771                f,
1772                "failed to publish connection file {}: {source}",
1773                path.display()
1774            ),
1775            Self::GenerateConnectionFile(err) => {
1776                write!(f, "failed to generate connection-file auth material: {err}")
1777            }
1778            Self::StartLockCreate { path, source } => {
1779                write!(
1780                    f,
1781                    "failed to create start lock {}: {source}",
1782                    path.display()
1783                )
1784            }
1785            Self::StartLockBusy { path, attempts } => write!(
1786                f,
1787                "start lock {} remained busy after {attempts} attempts",
1788                path.display()
1789            ),
1790            Self::RunDirLockCreate { path, source } => write!(
1791                f,
1792                "refusing to start: failed to lock run directory via {}: {source}",
1793                path.display()
1794            ),
1795            Self::RunDirBusy { path, holder_pid } => {
1796                write!(
1797                    f,
1798                    "refusing to start: run directory lock {} is held by another daemon",
1799                    path.display()
1800                )?;
1801                if let Some(pid) = holder_pid {
1802                    write!(f, " (pid {pid})")?;
1803                }
1804                write!(
1805                    f,
1806                    "; a live daemon owns this data home's run state (typically one started with a different XDG_RUNTIME_DIR)"
1807                )
1808            }
1809            Self::RemoveStale { path, source } => write!(
1810                f,
1811                "failed to remove stale connection file {}: {source}",
1812                path.display()
1813            ),
1814            Self::Bind { host, port, source } if source.kind() == io::ErrorKind::AddrInUse => {
1815                write!(
1816                    f,
1817                    "port {port} in use on loopback {host}: {source}; set the port in config"
1818                )
1819            }
1820            Self::Bind { host, port, source } => {
1821                write!(f, "failed to bind loopback TCP {host}:{port}: {source}")
1822            }
1823            Self::LocalAddr { host, source } => {
1824                write!(f, "failed to read local address for {host}: {source}")
1825            }
1826            Self::DaemonConfig(err) => write!(f, "failed to load daemon config: {err}"),
1827            Self::MachineId(err) => write!(f, "refusing to start: {err}"),
1828            Self::RunDir(err) => write!(f, "refusing to start: {err}"),
1829            Self::Serve(err) => write!(f, "daemon server failed: {err}"),
1830            Self::ServeJoin(err) => write!(f, "daemon server task failed: {err}"),
1831        }
1832    }
1833}
1834
1835impl Error for BootstrapError {
1836    fn source(&self) -> Option<&(dyn Error + 'static)> {
1837        match self {
1838            Self::StartupNotServing => None,
1839            #[cfg(unix)]
1840            Self::Signal(source) => Some(source),
1841            Self::InvalidPort { source, .. } => Some(source),
1842            Self::ConnectionFileRead { source, .. }
1843            | Self::ConnectionFileWrite { source, .. }
1844            | Self::GenerateConnectionFile(source) => Some(source),
1845            Self::StartLockCreate { source, .. }
1846            | Self::RunDirLockCreate { source, .. }
1847            | Self::RemoveStale { source, .. }
1848            | Self::Bind { source, .. }
1849            | Self::LocalAddr { source, .. } => Some(source),
1850            Self::DaemonConfig(err) => Some(err),
1851            Self::MachineId(err) => Some(err),
1852            Self::RunDir(err) => Some(err),
1853            Self::Serve(err) => Some(err),
1854            Self::ServeJoin(err) => Some(err),
1855            Self::StartLockBusy { .. } | Self::RunDirBusy { .. } => None,
1856        }
1857    }
1858}
1859
1860#[cfg(test)]
1861mod tests {
1862    mod activity {
1863        include!("activity_bootstrap_tests.rs");
1864    }
1865    mod startup_order {
1866        include!("startup_order_tests.rs");
1867    }
1868    use super::*;
1869    use crate::server::ServerAuth;
1870    use cortexkit_test_support::ScratchDir;
1871    #[cfg(target_os = "linux")]
1872    use std::collections::BTreeSet;
1873    use std::sync::Mutex;
1874    #[cfg(target_os = "linux")]
1875    use subc_control::ModuleProtocol;
1876    use subc_transport::MIN_KEY_LEN;
1877    use tokio::io::AsyncReadExt;
1878    use tokio::task::JoinHandle;
1879
1880    #[cfg(unix)]
1881    use std::os::unix::fs::PermissionsExt;
1882
1883    static ENV_LOCK: Mutex<()> = Mutex::new(());
1884
1885    #[test]
1886    fn normalized_build_provenance_preserves_real_values() {
1887        assert_eq!(normalized_build_provenance("abc"), Some("abc".to_string()));
1888    }
1889
1890    #[test]
1891    fn normalized_build_provenance_omits_unavailable_and_empty_values() {
1892        assert_eq!(normalized_build_provenance("unavailable"), None);
1893        assert_eq!(normalized_build_provenance(""), None);
1894    }
1895
1896    #[cfg(target_os = "linux")]
1897    fn current_cgroup_path_for_test() -> io::Result<PathBuf> {
1898        let cgroups = fs::read_to_string("/proc/self/cgroup")?;
1899        let relative = cgroups
1900            .lines()
1901            .find_map(|line| line.strip_prefix("0::"))
1902            .ok_or_else(|| {
1903                io::Error::new(io::ErrorKind::Unsupported, "cgroup v2 is unavailable")
1904            })?;
1905        Ok(Path::new("/sys/fs/cgroup").join(relative.trim_start_matches('/')))
1906    }
1907
1908    #[cfg(target_os = "linux")]
1909    fn module_cgroup_directories() -> io::Result<BTreeSet<OsString>> {
1910        let modules = current_cgroup_path_for_test()?.join("subc-modules");
1911        let entries = match fs::read_dir(modules) {
1912            Ok(entries) => entries,
1913            Err(error) if error.kind() == io::ErrorKind::NotFound => return Ok(BTreeSet::new()),
1914            Err(error) => return Err(error),
1915        };
1916        let mut directories = BTreeSet::new();
1917        for entry in entries {
1918            let entry = entry?;
1919            if entry.file_type()?.is_dir() {
1920                directories.insert(entry.file_name());
1921            }
1922        }
1923        Ok(directories)
1924    }
1925
1926    #[cfg(target_os = "linux")]
1927    async fn wait_for_path(path: &Path, task: &JoinHandle<Result<(), BootstrapError>>) {
1928        let deadline = tokio::time::Instant::now() + Duration::from_secs(5);
1929        while !path.exists() && tokio::time::Instant::now() < deadline {
1930            assert!(
1931                !task.is_finished(),
1932                "daemon exited before creating {}",
1933                path.display()
1934            );
1935            sleep(Duration::from_millis(10)).await;
1936        }
1937        assert!(path.exists(), "daemon did not create {}", path.display());
1938    }
1939
1940    /// An in-process daemon with the default (disabled) cgroup placement must
1941    /// leave the host's live module cgroup tree exactly as it found it.
1942    /// Red-checking this test (making it fail) creates a
1943    /// `cgroup-isolation-probe-*` directory in the live cgroup tree, which must be removed by hand.
1944    #[cfg(target_os = "linux")]
1945    #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
1946    async fn run_with_config_does_not_reconcile_the_ambient_cgroup_by_default() {
1947        let temp = unique_temp_dir("bootstrap-cgroup-default-disabled");
1948        let module_id = format!("cgroup-isolation-probe-{}", process::id());
1949        let before = module_cgroup_directories().expect("read ambient module cgroups before boot");
1950        assert!(
1951            !before.contains(&OsString::from(&module_id)),
1952            "isolation probe cgroup already exists before this daemon starts"
1953        );
1954        let capture = temp.join("logs").join(format!("{module_id}.stderr.log"));
1955        let module = ConfiguredModule {
1956            module_id,
1957            program: PathBuf::from("sh"),
1958            args: vec!["-c".to_string(), "sleep 30".to_string()],
1959            env: Vec::new(),
1960            log: None,
1961            enabled: true,
1962            reserved: false,
1963            reserved_prefixes: Vec::new(),
1964            protocol: ModuleProtocol::None,
1965            overlap: Default::default(),
1966            health: HealthConfig::default(),
1967            drain_timeout_ms: None,
1968            route_bind_relay_timeout_ms: None,
1969            restart: RestartPolicy::default(),
1970        };
1971        let config = BootstrapConfig::new(temp.join("connection.json"), 0)
1972            .with_configured_modules([module])
1973            .with_capture_logs_dir(temp.join("logs"))
1974            .with_terminal_journal_path(temp.join("terminals.jsonl"));
1975        let task = tokio::spawn(run_with_config(config));
1976
1977        wait_for_path(&capture, &task).await;
1978        let after = module_cgroup_directories().expect("read ambient module cgroups after boot");
1979
1980        task.abort();
1981        assert!(task
1982            .await
1983            .expect_err("aborted daemon task must cancel")
1984            .is_cancelled());
1985        assert_eq!(
1986            before, after,
1987            "an in-process daemon must not create or reconcile ambient module cgroups"
1988        );
1989    }
1990
1991    #[cfg(target_os = "linux")]
1992    #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
1993    async fn explicit_cgroup_root_is_prepared_inside_the_fixture_tree() {
1994        let temp = unique_temp_dir("bootstrap-cgroup-explicit-root");
1995        let cgroup_root = temp.join("cgroup");
1996        fs::create_dir(&cgroup_root).expect("create scratch cgroup root");
1997        fs::write(cgroup_root.join("cgroup.procs"), b"").expect("write scratch cgroup marker");
1998        let modules = cgroup_root.join("subc-modules");
1999        let config = BootstrapConfig::new(temp.join("connection.json"), 0)
2000            .with_cgroup_placement(CgroupPlacementConfig::Root(cgroup_root))
2001            .with_terminal_journal_path(temp.join("terminals.jsonl"));
2002        let task = tokio::spawn(run_with_config(config));
2003
2004        wait_for_path(&modules, &task).await;
2005
2006        task.abort();
2007        assert!(task
2008            .await
2009            .expect_err("aborted daemon task must cancel")
2010            .is_cancelled());
2011        assert!(
2012            fs::read_dir(&modules)
2013                .expect("read prepared modules directory")
2014                .next()
2015                .is_none(),
2016            "the delegation probe must clean up after itself"
2017        );
2018    }
2019
2020    struct EnvGuard {
2021        key: &'static str,
2022        previous: Option<OsString>,
2023    }
2024
2025    impl EnvGuard {
2026        fn set(key: &'static str, value: &Path) -> Self {
2027            let previous = env::var_os(key);
2028            env::set_var(key, value);
2029            Self { key, previous }
2030        }
2031
2032        fn set_str(key: &'static str, value: &str) -> Self {
2033            let previous = env::var_os(key);
2034            env::set_var(key, value);
2035            Self { key, previous }
2036        }
2037
2038        fn unset(key: &'static str) -> Self {
2039            let previous = env::var_os(key);
2040            env::remove_var(key);
2041            Self { key, previous }
2042        }
2043    }
2044
2045    impl Drop for EnvGuard {
2046        fn drop(&mut self) {
2047            match &self.previous {
2048                Some(value) => env::set_var(self.key, value),
2049                None => env::remove_var(self.key),
2050            }
2051        }
2052    }
2053
2054    fn unique_temp_dir(name: &str) -> ScratchDir {
2055        ScratchDir::new(name)
2056    }
2057
2058    fn temp_connection_file_path(name: &str) -> (ScratchDir, PathBuf) {
2059        let dir = unique_temp_dir(name);
2060        let path = dir.join("conn.json");
2061        (dir, path)
2062    }
2063
2064    fn auth_for(info: &ConnectionInfo) -> ServerAuth {
2065        ServerAuth::new(info.key.clone(), info.daemon_id, info.daemon_ver.clone())
2066    }
2067
2068    fn start_server(bound: BoundDaemon) -> JoinHandle<Result<(), ServerError>> {
2069        let auth = auth_for(&bound.connection_info);
2070        tokio::spawn(serve_listeners(
2071            bound.listeners,
2072            Arc::new(Router::with_default_self_handler()),
2073            auth,
2074        ))
2075    }
2076
2077    fn expect_bound(outcome: Outcome) -> BoundDaemon {
2078        match outcome {
2079            Outcome::Bound(bound) => bound,
2080            Outcome::AlreadyRunning => panic!("fresh connection file unexpectedly had a daemon"),
2081        }
2082    }
2083
2084    async fn connect_from_info(conn: &ConnectionInfo) -> io::Result<TcpStream> {
2085        let endpoint = conn
2086            .endpoints
2087            .first()
2088            .expect("test connection file should have an endpoint");
2089        let ip: IpAddr = endpoint.host.parse().unwrap();
2090        TcpStream::connect(SocketAddr::new(ip, endpoint.port)).await
2091    }
2092
2093    fn make_connection_info(port: u16) -> ConnectionInfo {
2094        ConnectionInfo {
2095            schema: SCHEMA_VERSION,
2096            wire_version: Some(PROTOCOL_VERSION),
2097            endpoints: vec![Endpoint {
2098                host: "127.0.0.1".to_owned(),
2099                port,
2100            }],
2101            key: generate_key().unwrap(),
2102            daemon_id: generate_daemon_id().unwrap(),
2103            pid: process::id(),
2104            daemon_ver: "test-subc".to_owned(),
2105        }
2106    }
2107
2108    fn write_raw_owner_only_connection_file(path: &Path, contents: &[u8]) {
2109        fs::write(path, contents).unwrap();
2110        #[cfg(unix)]
2111        fs::set_permissions(path, fs::Permissions::from_mode(0o600)).unwrap();
2112    }
2113
2114    fn assert_owner_only_connection_file(path: &Path) {
2115        // `path` is only inspected on Unix (mode bits); on Windows the owner-only
2116        // guarantee comes from the inherited %TEMP% ACL, nothing to assert here.
2117        #[cfg(unix)]
2118        {
2119            let mode = fs::metadata(path).unwrap().permissions().mode() & 0o777;
2120            assert_eq!(mode, 0o600);
2121        }
2122        #[cfg(not(unix))]
2123        let _ = path;
2124    }
2125
2126    #[test]
2127    fn connection_file_path_uses_xdg_runtime_dir_when_set() {
2128        let _env_lock = ENV_LOCK.lock().unwrap();
2129        let runtime_dir = unique_temp_dir("xdg-runtime");
2130        let _xdg = EnvGuard::set("XDG_RUNTIME_DIR", runtime_dir.path());
2131
2132        assert_eq!(
2133            connection_file_path(),
2134            runtime_dir.join(CONNECTION_FILE_NAME)
2135        );
2136    }
2137
2138    #[test]
2139    fn connection_file_path_source_is_xdg_runtime_dir_when_set() {
2140        let runtime_dir = OsString::from("/run/user/1000");
2141
2142        let (path, source) = connection_file_path_with_source(Some(runtime_dir));
2143
2144        assert_eq!(
2145            path,
2146            PathBuf::from("/run/user/1000").join(CONNECTION_FILE_NAME)
2147        );
2148        assert_eq!(source, ConnectionFileSource::XdgRuntimeDir);
2149    }
2150
2151    #[test]
2152    fn connection_file_path_falls_back_to_temp_dir_with_user_token_when_xdg_unset() {
2153        let _env_lock = ENV_LOCK.lock().unwrap();
2154        let _xdg = EnvGuard::unset("XDG_RUNTIME_DIR");
2155
2156        assert_eq!(
2157            connection_file_path(),
2158            env::temp_dir().join(format!("subc-{}.connection.json", user_connection_token()))
2159        );
2160    }
2161
2162    #[test]
2163    fn connection_file_path_source_is_temp_dir_when_xdg_unset() {
2164        let (path, source) = connection_file_path_with_source(None);
2165
2166        assert_eq!(
2167            path,
2168            env::temp_dir().join(format!("subc-{}.connection.json", user_connection_token()))
2169        );
2170        assert_eq!(source, ConnectionFileSource::TempDirFallback);
2171    }
2172
2173    /// Concurrent callers must derive one token. The former temp-file uid probe
2174    /// could fail transiently (same-tick name collision, fd exhaustion) and send
2175    /// the loser down the env-derived fallback with a different identity for the
2176    /// same user; this fence keeps identity independent of filesystem luck.
2177    #[test]
2178    fn user_connection_token_is_stable_under_concurrent_callers() {
2179        let expected = user_connection_token();
2180        let workers: Vec<_> = (0..32)
2181            .map(|_| {
2182                std::thread::spawn(|| (0..40).map(|_| user_connection_token()).collect::<Vec<_>>())
2183            })
2184            .collect();
2185        for worker in workers {
2186            for token in worker.join().expect("probe thread") {
2187                assert_eq!(token, expected, "token diverged under concurrent probes");
2188            }
2189        }
2190    }
2191
2192    #[test]
2193    fn connection_file_path_source_is_temp_dir_when_xdg_empty() {
2194        let (path, source) = connection_file_path_with_source(Some(OsString::new()));
2195
2196        assert_eq!(
2197            path,
2198            env::temp_dir().join(format!("subc-{}.connection.json", user_connection_token()))
2199        );
2200        assert_eq!(source, ConnectionFileSource::TempDirFallback);
2201    }
2202
2203    /// The shipped binary is the only caller that journals terminal exits into
2204    /// the real run directory, so its constructor must supply that path itself;
2205    /// the in-process default leaves it unset.
2206    #[test]
2207    fn daemon_binary_config_journals_and_captures_into_the_run_dir() {
2208        let _env_lock = ENV_LOCK.lock().unwrap();
2209        let root = unique_temp_dir("daemon-binary-config");
2210        let data_home = root.join("data");
2211        let _data = EnvGuard::set("XDG_DATA_HOME", &data_home);
2212        let _config = EnvGuard::set("XDG_CONFIG_HOME", &root.join("config"));
2213        let _port = EnvGuard::unset(SUBC_PORT_ENV);
2214
2215        let config = BootstrapConfig::from_env_for_daemon_binary().unwrap();
2216
2217        let run_dir = data_home.join("cortexkit").join("run");
2218        assert_eq!(
2219            config.terminal_journal_path,
2220            Some(run_dir.join("terminals.jsonl"))
2221        );
2222        assert_eq!(config.capture_logs_dir, Some(run_dir.join("logs")));
2223        assert_eq!(
2224            BootstrapConfig::new(root.join("connection.json"), 0).terminal_journal_path,
2225            None,
2226            "an in-process config must not journal anywhere unless asked to"
2227        );
2228    }
2229
2230    #[test]
2231    fn daemon_binary_config_refuses_a_relative_data_home() {
2232        let _env_lock = ENV_LOCK.lock().unwrap();
2233        let root = unique_temp_dir("daemon-binary-relative-data");
2234        let _data = EnvGuard::set_str("XDG_DATA_HOME", "relative-data-home");
2235        let _config = EnvGuard::set("XDG_CONFIG_HOME", &root.join("config"));
2236        let _port = EnvGuard::unset(SUBC_PORT_ENV);
2237
2238        let error = BootstrapConfig::from_env_for_daemon_binary()
2239            .expect_err("a relative data home must refuse the daemon binary's config");
2240        assert!(
2241            matches!(error, BootstrapError::RunDir(_)),
2242            "expected a run-directory refusal, got {error}"
2243        );
2244        assert!(error.to_string().contains("XDG_DATA_HOME"), "{error}");
2245    }
2246
2247    #[test]
2248    fn privileged_operator_prompt_is_loaded_by_both_bootstrap_config_paths() {
2249        let _env_lock = ENV_LOCK.lock().unwrap();
2250        let _port = EnvGuard::unset(SUBC_PORT_ENV);
2251        let (_dir, conn_path) = temp_connection_file_path("privileged-operator-prompt");
2252        let config_path = conn_path.with_file_name("subc.jsonc");
2253        for (doc, expected) in [
2254            (None, daemon_config::PrivilegedOperatorPrompt::Log),
2255            (
2256                Some(r#"{"version":1}"#),
2257                daemon_config::PrivilegedOperatorPrompt::Log,
2258            ),
2259            (
2260                Some(r#"{"version":1,"privileged_operator_prompt":"log"}"#),
2261                daemon_config::PrivilegedOperatorPrompt::Log,
2262            ),
2263            (
2264                Some(r#"{"version":1,"privileged_operator_prompt":"enforce"}"#),
2265                daemon_config::PrivilegedOperatorPrompt::Enforce,
2266            ),
2267        ] {
2268            if let Some(doc) = doc {
2269                fs::write(&config_path, doc).unwrap();
2270            }
2271            assert_eq!(
2272                BootstrapConfig::from_env_with_daemon_config_path(&config_path)
2273                    .unwrap()
2274                    .privileged_operator_prompt,
2275                expected
2276            );
2277            assert_eq!(
2278                BootstrapConfig::new(&conn_path, 0)
2279                    .with_daemon_config_path(&config_path)
2280                    .unwrap()
2281                    .privileged_operator_prompt,
2282                expected
2283            );
2284        }
2285    }
2286
2287    #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
2288    async fn privileged_operator_prompt_is_snapshotted_by_running_daemon() {
2289        let dir = unique_temp_dir("privileged-operator-prompt-bootstrap");
2290        let connection_path = dir.join("connection.json");
2291        let config_path = dir.join("subc.jsonc");
2292        fs::write(
2293            &config_path,
2294            r#"{"version":1,"privileged_operator_prompt":"enforce"}"#,
2295        )
2296        .unwrap();
2297        let config = BootstrapConfig::new(&connection_path, 0)
2298            .with_daemon_config_path(&config_path)
2299            .unwrap();
2300        let daemon = AbortOnDrop::new(tokio::spawn(run_with_config(config)));
2301        tokio::time::timeout(Duration::from_secs(5), async {
2302            while !connection_path.exists() {
2303                assert!(
2304                    !daemon.handle.is_finished(),
2305                    "daemon exited before publishing"
2306                );
2307                tokio::task::yield_now().await;
2308            }
2309        })
2310        .await
2311        .unwrap();
2312        let info = connection_file::read(&connection_path).unwrap();
2313        let mut client = connect_from_info(&info).await.unwrap();
2314        authenticate_client(&mut client, &info, Duration::from_secs(5))
2315            .await
2316            .unwrap();
2317        for (correlation, mode, expected) in [
2318            (1, "enforce", Vec::new()),
2319            (2, "log", vec!["privileged_operator_prompt"]),
2320            (3, "log", vec!["privileged_operator_prompt"]),
2321        ] {
2322            fs::write(
2323                &config_path,
2324                serde_json::json!({"version":1,"privileged_operator_prompt":mode}).to_string(),
2325            )
2326            .unwrap();
2327            let frame = crate::Frame::build(
2328                subc_protocol::FrameType::Request,
2329                subc_protocol::Flags::new(false, subc_protocol::Priority::Interactive, false),
2330                0,
2331                0,
2332                correlation,
2333                serde_json::to_vec(&subc_control::ClientControlRequest::SupervisorRescan {
2334                    preview: true,
2335                })
2336                .unwrap(),
2337            )
2338            .unwrap();
2339            crate::write_frame(&mut client, &frame).await.unwrap();
2340            let reply =
2341                tokio::time::timeout(Duration::from_secs(5), crate::read_frame(&mut client))
2342                    .await
2343                    .unwrap()
2344                    .unwrap()
2345                    .unwrap();
2346            assert_eq!(reply.header.ty, subc_protocol::FrameType::Response);
2347            assert_eq!(reply.header.corr, correlation);
2348            let subc_control::ClientControlResponse::SupervisorRescan { result } =
2349                serde_json::from_slice(&reply.body).unwrap()
2350            else {
2351                panic!("expected a rescan response");
2352            };
2353            assert_eq!(result.restart_required, expected);
2354        }
2355    }
2356
2357    #[test]
2358    fn configured_port_uses_default_config_and_env_override() {
2359        let _env_lock = ENV_LOCK.lock().unwrap();
2360        let (_dir, conn_path) = temp_connection_file_path("daemon-config-port");
2361        let config_path = conn_path.with_file_name("subc.jsonc");
2362
2363        let _port = EnvGuard::unset(SUBC_PORT_ENV);
2364        assert_eq!(
2365            BootstrapConfig::from_env_with_daemon_config_path(&config_path)
2366                .unwrap()
2367                .port,
2368            DEFAULT_SUBC_PORT
2369        );
2370
2371        fs::write(&config_path, r#"{ "version": 1, "port": 8123 }"#).unwrap();
2372        assert_eq!(
2373            BootstrapConfig::from_env_with_daemon_config_path(&config_path)
2374                .unwrap()
2375                .port,
2376            8123
2377        );
2378
2379        let _port = EnvGuard::set_str(SUBC_PORT_ENV, "9012");
2380        assert_eq!(
2381            BootstrapConfig::from_env_with_daemon_config_path(&config_path)
2382                .unwrap()
2383                .port,
2384            9012
2385        );
2386    }
2387
2388    #[tokio::test]
2389    async fn second_singleton_probe_against_served_tcp_daemon_reports_already_running() {
2390        let (_dir, path) = temp_connection_file_path("already-running");
2391
2392        let bound = expect_bound(ensure_singleton(&path, 0).await.unwrap());
2393        let server = start_server(bound);
2394
2395        let second = ensure_singleton(&path, 0).await.unwrap();
2396        assert!(matches!(second, Outcome::AlreadyRunning));
2397
2398        server.abort();
2399        let _ = server.await;
2400    }
2401
2402    #[tokio::test]
2403    async fn daemon_connection_file_publishes_protocol_wire_version() {
2404        let (_dir, path) = temp_connection_file_path("wire-version");
2405        let bound = expect_bound(ensure_singleton(&path, 0).await.unwrap());
2406        assert_eq!(bound.connection_info.wire_version, Some(PROTOCOL_VERSION));
2407        assert_eq!(
2408            connection_file::read(&path).unwrap().wire_version,
2409            Some(PROTOCOL_VERSION)
2410        );
2411
2412        drop(bound.listeners);
2413    }
2414
2415    #[tokio::test]
2416    async fn stale_unbound_connection_file_is_reclaimed() {
2417        let (_dir, path) = temp_connection_file_path("stale-reclaim");
2418        let stale = TcpListener::bind((Ipv4Addr::LOCALHOST, 0)).await.unwrap();
2419        let stale_port = stale.local_addr().unwrap().port();
2420        drop(stale);
2421        let stale_info = make_connection_info(stale_port);
2422        write_atomic(&path, &stale_info).unwrap();
2423
2424        let bound = expect_bound(ensure_singleton(&path, 0).await.unwrap());
2425        assert_ne!(bound.connection_info.key, stale_info.key);
2426        drop(bound.listeners);
2427    }
2428
2429    #[cfg(unix)]
2430    #[tokio::test]
2431    async fn ensure_singleton_reclaims_insecure_connection_file() {
2432        let (_dir, path) = temp_connection_file_path("insecure-reclaim");
2433        let stale = TcpListener::bind((Ipv4Addr::LOCALHOST, 0)).await.unwrap();
2434        let stale_port = stale.local_addr().unwrap().port();
2435        drop(stale);
2436        let stale_info = make_connection_info(stale_port);
2437        write_atomic(&path, &stale_info).unwrap();
2438        fs::set_permissions(&path, fs::Permissions::from_mode(0o644)).unwrap();
2439
2440        let bound = expect_bound(ensure_singleton(&path, 0).await.unwrap());
2441        assert_ne!(bound.connection_info.key, stale_info.key);
2442        assert_ne!(bound.connection_info.daemon_id, stale_info.daemon_id);
2443        assert_owner_only_connection_file(&path);
2444
2445        drop(bound.listeners);
2446    }
2447
2448    #[tokio::test]
2449    async fn ensure_singleton_reclaims_non_loopback_connection_file() {
2450        let (_dir, path) = temp_connection_file_path("non-loopback-reclaim");
2451        let mut stale_info = make_connection_info(8757);
2452        stale_info.endpoints = vec![Endpoint {
2453            host: "192.0.2.10".to_owned(),
2454            port: 8757,
2455        }];
2456        write_atomic(&path, &stale_info).unwrap();
2457
2458        let bound = expect_bound(ensure_singleton(&path, 0).await.unwrap());
2459        assert_ne!(bound.connection_info.key, stale_info.key);
2460        assert_ne!(bound.connection_info.daemon_id, stale_info.daemon_id);
2461        assert!(bound
2462            .connection_info
2463            .endpoints
2464            .iter()
2465            .all(|endpoint| endpoint.host.parse::<IpAddr>().unwrap().is_loopback()));
2466        assert_owner_only_connection_file(&path);
2467
2468        drop(bound.listeners);
2469    }
2470
2471    #[tokio::test]
2472    async fn ensure_singleton_reclaims_invalid_connection_file_shapes() {
2473        let mut unsupported_schema = make_connection_info(8757);
2474        unsupported_schema.schema = SCHEMA_VERSION + 1;
2475
2476        let mut empty_endpoints = make_connection_info(8757);
2477        empty_endpoints.endpoints.clear();
2478
2479        let mut short_key = make_connection_info(8757);
2480        short_key.key = vec![0x5A; MIN_KEY_LEN - 1];
2481
2482        let cases = vec![
2483            (
2484                "unsupported-schema",
2485                serde_json::to_vec(&unsupported_schema).unwrap(),
2486                Some(unsupported_schema),
2487            ),
2488            (
2489                "empty-endpoints",
2490                serde_json::to_vec(&empty_endpoints).unwrap(),
2491                Some(empty_endpoints),
2492            ),
2493            (
2494                "short-key",
2495                serde_json::to_vec(&short_key).unwrap(),
2496                Some(short_key),
2497            ),
2498            ("invalid-json", b"{not valid connection json".to_vec(), None),
2499        ];
2500
2501        for (label, contents, old_info) in cases {
2502            let (_dir, path) = temp_connection_file_path(label);
2503            write_raw_owner_only_connection_file(&path, &contents);
2504
2505            let bound = expect_bound(ensure_singleton(&path, 0).await.unwrap());
2506            if let Some(old_info) = old_info {
2507                assert_ne!(bound.connection_info.key, old_info.key, "{label}");
2508                assert_ne!(
2509                    bound.connection_info.daemon_id, old_info.daemon_id,
2510                    "{label}"
2511                );
2512            }
2513            assert!(bound.connection_info.key.len() >= MIN_KEY_LEN, "{label}");
2514            assert_ne!(bound.connection_info.daemon_id, [0u8; 16], "{label}");
2515            assert_owner_only_connection_file(&path);
2516
2517            drop(bound.listeners);
2518        }
2519    }
2520
2521    #[tokio::test]
2522    async fn foreign_reused_port_connection_file_is_reclaimed_after_auth_probe_fails() {
2523        let (_dir, path) = temp_connection_file_path("foreign-reclaim");
2524        let foreign = TcpListener::bind((Ipv4Addr::LOCALHOST, 0)).await.unwrap();
2525        let foreign_port = foreign.local_addr().unwrap().port();
2526        write_atomic(&path, &make_connection_info(foreign_port)).unwrap();
2527        let foreign_task = tokio::spawn(async move {
2528            if let Ok((mut stream, _)) = foreign.accept().await {
2529                let mut buf = [0u8; 64];
2530                let _ = stream.read(&mut buf).await;
2531            }
2532        });
2533
2534        let bound = expect_bound(ensure_singleton(&path, 0).await.unwrap());
2535        assert!(bound
2536            .connection_info
2537            .endpoints
2538            .iter()
2539            .all(|endpoint| endpoint.port != foreign_port));
2540
2541        drop(bound.listeners);
2542        let _ = foreign_task.await;
2543    }
2544
2545    #[tokio::test]
2546    async fn stale_start_lock_file_is_reclaimable() {
2547        let (_dir, path) = temp_connection_file_path("start-lock-stale-file");
2548        let lock_path = start_lock_path(&path);
2549        drop(open_owner_only_lock(&lock_path).unwrap());
2550        assert!(lock_path.is_file());
2551
2552        let lock = StartLock::acquire(&path).await.unwrap();
2553        assert!(lock_path.is_file());
2554
2555        drop(lock);
2556        assert!(lock_path.is_file());
2557    }
2558
2559    #[tokio::test]
2560    async fn held_start_lock_blocks_second_acquire_until_release() {
2561        let (_dir, path) = temp_connection_file_path("start-lock-held");
2562        let lock_path = start_lock_path(&path);
2563        let first = StartLock::acquire(&path).await.unwrap();
2564
2565        let err = match StartLock::acquire(&path).await {
2566            Ok(_) => panic!("second acquire while held must stay busy"),
2567            Err(err) => err,
2568        };
2569        assert!(matches!(
2570            err,
2571            BootstrapError::StartLockBusy {
2572                ref path,
2573                attempts: START_LOCK_RETRIES,
2574            } if path == &lock_path
2575        ));
2576
2577        drop(first);
2578
2579        let second = StartLock::acquire(&path)
2580            .await
2581            .expect("released advisory lock should be reclaimable");
2582        drop(second);
2583    }
2584
2585    #[tokio::test]
2586    async fn bind_conflict_on_fixed_port_fails_loud_without_reselecting() {
2587        let (_dir, path) = temp_connection_file_path("bind-conflict");
2588        let occupied = TcpListener::bind((Ipv4Addr::LOCALHOST, 0)).await.unwrap();
2589        let occupied_port = occupied.local_addr().unwrap().port();
2590
2591        let err = ensure_singleton(&path, occupied_port).await.unwrap_err();
2592        assert!(matches!(
2593            err,
2594            BootstrapError::Bind { ref source, .. } if source.kind() == io::ErrorKind::AddrInUse
2595        ));
2596        assert!(err.to_string().contains("set the port in config"));
2597
2598        drop(occupied);
2599    }
2600
2601    #[tokio::test]
2602    async fn ephemeral_ipv6_collision_still_serves_ipv4() {
2603        let (listeners, endpoints) = bind_loopback_with(0, |port| async move {
2604            let occupied = TcpListener::bind((Ipv6Addr::LOCALHOST, port)).await?;
2605            let result = TcpListener::bind((Ipv6Addr::LOCALHOST, port)).await;
2606            drop(occupied);
2607            result
2608        })
2609        .await
2610        .expect("an incidental IPv6 collision must not prevent an ephemeral daemon from starting");
2611        assert_eq!(endpoints.len(), 1);
2612        assert_eq!(endpoints[0].host, "127.0.0.1");
2613        TcpStream::connect((Ipv4Addr::LOCALHOST, endpoints[0].port))
2614            .await
2615            .unwrap();
2616        drop(listeners);
2617    }
2618
2619    #[tokio::test]
2620    async fn unpublished_boot_holds_singleton_lock_until_publication() {
2621        let dir = ScratchDir::new("unpublished-singleton-lock");
2622        let path = dir.join("connection.json");
2623        let bound = expect_bound(
2624            ensure_singleton_inner(BootstrapConfig::new(&path, 0), false)
2625                .await
2626                .unwrap(),
2627        );
2628        assert!(!path.exists());
2629        let contender = open_owner_only_lock(&start_lock_path(&path)).unwrap();
2630        assert!(
2631            matches!(FileExt::try_lock(&contender), Err(TryLockError::WouldBlock)),
2632            "an unpublished boot must keep its singleton claim"
2633        );
2634        drop(bound);
2635        FileExt::try_lock(&contender).unwrap();
2636    }
2637
2638    #[tokio::test]
2639    async fn start_lock_drop_releases_claim_with_a_shared_file_description() {
2640        let dir = ScratchDir::new("shared-singleton-lock");
2641        let path = dir.join("connection.json");
2642        let bound = expect_bound(
2643            ensure_singleton_inner(BootstrapConfig::new(&path, 0), false)
2644                .await
2645                .unwrap(),
2646        );
2647        // try_clone shares the open file description just as fork does,
2648        // without relying on the timing of another test's child spawn.
2649        let inherited = bound
2650            .publication_lock
2651            .as_ref()
2652            .unwrap()
2653            ._file
2654            .try_clone()
2655            .unwrap();
2656        let contender = open_owner_only_lock(&start_lock_path(&path)).unwrap();
2657        assert!(matches!(
2658            FileExt::try_lock(&contender),
2659            Err(TryLockError::WouldBlock)
2660        ));
2661        drop(bound);
2662        FileExt::try_lock(&contender).expect("dropping bootstrap releases even an inherited lock");
2663        drop(inherited);
2664    }
2665
2666    #[tokio::test]
2667    async fn readiness_probe_requires_an_authenticating_server() {
2668        let dir = ScratchDir::new("unserved-readiness");
2669        let bound = expect_bound(
2670            ensure_singleton_inner(BootstrapConfig::new(dir.join("connection.json"), 0), false)
2671                .await
2672                .unwrap(),
2673        );
2674        assert!(matches!(
2675            verify_serving(&bound.connection_info).await,
2676            Err(BootstrapError::StartupNotServing)
2677        ));
2678    }
2679
2680    #[cfg(unix)]
2681    #[tokio::test]
2682    async fn orphan_sweep_finishes_before_connection_file_publication() {
2683        use crate::live_children::{ExecutableIdentity, LiveChild};
2684        let dir = ScratchDir::new("publish-after-sweep");
2685        let path = dir.join("connection.json");
2686        let record = dir.join("live-children.json");
2687        let ready = dir.join("ready");
2688        let mut child = std::process::Command::new("/bin/sh")
2689            .args([
2690                "-c",
2691                "trap '' TERM; : > \"$0\"; read -r _",
2692                ready.to_str().unwrap(),
2693            ])
2694            .stdin(std::process::Stdio::piped())
2695            .env("XDG_DATA_HOME", dir.path())
2696            .env("XDG_RUNTIME_DIR", dir.path())
2697            .env("XDG_CONFIG_HOME", dir.path())
2698            .spawn()
2699            .unwrap();
2700        timeout(Duration::from_secs(5), async {
2701            while !ready.exists() {
2702                sleep(Duration::from_millis(5)).await;
2703            }
2704        })
2705        .await
2706        .unwrap();
2707        let observed = subc_os::Process::open(child.id())
2708            .unwrap()
2709            .unwrap()
2710            .observe()
2711            .unwrap();
2712        crate::live_children::write_record(
2713            &record,
2714            &[LiveChild {
2715                module_id: "old-sleep".into(),
2716                pid: child.id(),
2717                protocol: subc_control::ModuleProtocol::None,
2718                start_time: Some(observed.start_time),
2719                executable: observed.executable.map(ExecutableIdentity::from),
2720                cgroup_name: None,
2721            }],
2722        )
2723        .unwrap();
2724        let config = BootstrapConfig::new(&path, 0).with_live_children_record(&record);
2725        let daemon = tokio::spawn(run_with_config(config));
2726        timeout(Duration::from_secs(5), async {
2727            while !path.exists() {
2728                sleep(Duration::from_millis(5)).await;
2729            }
2730        })
2731        .await
2732        .unwrap();
2733        assert!(path.exists());
2734        let ended = child.try_wait().unwrap().is_some();
2735        if !ended {
2736            child.kill().unwrap();
2737        }
2738        child.wait().unwrap();
2739        if ended {
2740            assert!(matches!(probe_existing(&path).await.unwrap(), Probe::Live));
2741        }
2742        daemon.abort();
2743        let _ = daemon.await;
2744        assert!(
2745            ended,
2746            "the published daemon must not still owe its orphan sweep"
2747        );
2748    }
2749
2750    /// Publishing the connection file fsyncs, which can take seconds on a busy
2751    /// disk, so the daemon must do it off the async worker thread. The hook
2752    /// blocks inside the write until another task on the same single-threaded
2753    /// runtime has run. Writing on the worker thread would starve that task: the
2754    /// hook then gives up after its deadline and the test fails, as it does if
2755    /// startup publishes without going through the hooked write at all.
2756    #[tokio::test(flavor = "current_thread")]
2757    async fn connection_file_publication_leaves_the_async_runtime_free() {
2758        use std::sync::atomic::{AtomicBool, Ordering};
2759
2760        let (_dir, path) = temp_connection_file_path("publish-off-runtime");
2761        let write_started = Arc::new(AtomicBool::new(false));
2762        let (progress_tx, progress_rx) = std::sync::mpsc::channel::<()>();
2763        let (verdict_tx, verdict_rx) = std::sync::mpsc::channel::<bool>();
2764        let hook_started = Arc::clone(&write_started);
2765        connection_file_write_hook::install(&path, move || {
2766            hook_started.store(true, Ordering::SeqCst);
2767            let progressed = progress_rx.recv_timeout(Duration::from_secs(5)).is_ok();
2768            let _ = verdict_tx.send(progressed);
2769        });
2770        // Signals only after the write has started, so a message can only come
2771        // from this task running while the write is still blocked.
2772        let ticker = tokio::spawn(async move {
2773            loop {
2774                if write_started.load(Ordering::SeqCst) {
2775                    let _ = progress_tx.send(());
2776                    return;
2777                }
2778                sleep(Duration::from_millis(5)).await;
2779            }
2780        });
2781
2782        let daemon = tokio::spawn(run_with_config(BootstrapConfig::new(&path, 0)));
2783        let published = timeout(Duration::from_secs(15), async {
2784            while !path.exists() && !daemon.is_finished() {
2785                sleep(Duration::from_millis(5)).await;
2786            }
2787        })
2788        .await;
2789        let daemon_exited = daemon.is_finished();
2790        daemon.abort();
2791        let daemon_result = daemon.await;
2792        ticker.abort();
2793
2794        assert!(
2795            published.is_ok() && !daemon_exited,
2796            "the daemon must start and publish its connection file: {daemon_result:?}"
2797        );
2798        match verdict_rx.try_recv() {
2799            Ok(true) => {}
2800            Ok(false) => panic!(
2801                "no other task on the runtime ran while the connection file was being written: the write is blocking the async worker thread"
2802            ),
2803            Err(_) => panic!("startup published the connection file without the hooked write"),
2804        }
2805    }
2806
2807    /// A write that never reports back must stop startup with an error naming
2808    /// the connection file, never be taken as published.
2809    #[tokio::test]
2810    async fn a_panicking_connection_file_write_fails_startup() {
2811        let (_dir, path) = temp_connection_file_path("publish-panics");
2812        connection_file_write_hook::install(&path, || {
2813            panic!("deliberate panic inside the connection-file write")
2814        });
2815
2816        let error = timeout(
2817            Duration::from_secs(15),
2818            run_with_config(BootstrapConfig::new(&path, 0)),
2819        )
2820        .await
2821        .expect("startup must end once the write task fails")
2822        .expect_err("startup must fail when the connection-file write panics");
2823
2824        assert!(
2825            matches!(&error, BootstrapError::ConnectionFileWrite { path: failed, .. } if failed == &path),
2826            "expected a connection-file write error, got {error:?}"
2827        );
2828        let message = error.to_string();
2829        assert!(message.contains(&path.display().to_string()), "{message}");
2830        assert!(message.contains("panicked"), "{message}");
2831        assert!(!path.exists(), "no connection file may be published");
2832    }
2833
2834    #[cfg(target_os = "macos")]
2835    #[test]
2836    fn nofile_raise_worker() {
2837        let Some(path) = std::env::var_os("SUBC_NOFILE_TEST_FILE") else {
2838            return;
2839        };
2840        let (_, hard) = rlimit::Resource::NOFILE.get().unwrap();
2841        rlimit::Resource::NOFILE.set(256, hard).unwrap();
2842        raise_nofile_limit_with_ceiling(Some(10240));
2843        std::fs::write(path, rlimit::Resource::NOFILE.get().unwrap().0.to_string()).unwrap();
2844    }
2845
2846    #[cfg(target_os = "macos")]
2847    #[test]
2848    fn macos_nofile_raise_clamps_to_kernel_ceiling() {
2849        let dir = ScratchDir::new("nofile-ceiling");
2850        let result = dir.join("soft");
2851        let status = std::process::Command::new(std::env::current_exe().unwrap())
2852            .args(["--exact", "bootstrap::tests::nofile_raise_worker"])
2853            .env("SUBC_NOFILE_TEST_FILE", &result)
2854            .env("XDG_DATA_HOME", dir.path())
2855            .env("XDG_RUNTIME_DIR", dir.path())
2856            .env("XDG_CONFIG_HOME", dir.path())
2857            .status()
2858            .unwrap();
2859        assert!(status.success());
2860        let soft: u64 = std::fs::read_to_string(result).unwrap().parse().unwrap();
2861        assert!(
2862            soft > 256 && soft <= 10240,
2863            "the inherited limit must be raised within the kernel ceiling; got {soft}"
2864        );
2865    }
2866
2867    #[tokio::test]
2868    async fn key_rotation_republishes_new_material_and_old_file_fails_auth() {
2869        let (_dir, path) = temp_connection_file_path("key-rotation");
2870        let first = expect_bound(ensure_singleton(&path, 0).await.unwrap());
2871        let old_info = first.connection_info.clone();
2872        let fixed_port = old_info.endpoints[0].port;
2873        drop(first.listeners);
2874
2875        let second = expect_bound(ensure_singleton(&path, fixed_port).await.unwrap());
2876        let new_info = second.connection_info.clone();
2877        assert_ne!(old_info.key, new_info.key);
2878        assert_ne!(old_info.daemon_id, new_info.daemon_id);
2879        let server = start_server(second);
2880
2881        let mut old_stream = connect_from_info(&old_info).await.unwrap();
2882        let old_auth = authenticate_client(&mut old_stream, &old_info, PROBE_AUTH_DEADLINE).await;
2883        assert!(
2884            old_auth.is_err(),
2885            "old key must not authenticate after restart"
2886        );
2887
2888        let reread = connection_file::read(&path).unwrap();
2889        let mut new_stream = connect_from_info(&reread).await.unwrap();
2890        authenticate_client(&mut new_stream, &reread, PROBE_AUTH_DEADLINE)
2891            .await
2892            .unwrap();
2893
2894        server.abort();
2895        let _ = server.await;
2896    }
2897
2898    #[cfg(unix)]
2899    #[tokio::test]
2900    async fn published_connection_file_permissions_are_owner_only() {
2901        let (_dir, path) = temp_connection_file_path("permissions");
2902        let bound = expect_bound(ensure_singleton(&path, 0).await.unwrap());
2903
2904        let mode = fs::metadata(&path).unwrap().permissions().mode() & 0o777;
2905        assert_eq!(mode, 0o600);
2906
2907        drop(bound.listeners);
2908    }
2909}