scv_server/
daemon.rs

1//! The daemon: the socket listener and its lock, supervision of components
2//! and delegated runs, and shutdown. [`run_stdio`] serves one connection
3//! without a daemon.
4
5use std::{path::Path, sync::Arc, time::Duration};
6
7use anyhow::{Context, Result, anyhow};
8use scv_client::Layout;
9use scv_tools::delegation::{self as delegations, DelegationRegistry};
10use tokio::{
11    net::{UnixListener, UnixStream},
12    sync::Mutex,
13};
14use tokio_util::{sync::CancellationToken, task::TaskTracker};
15
16use crate::{
17    components,
18    config::{ConfigOverrides, Instance},
19    confirm,
20    connection::{Shared, run_managed},
21    disk, restart,
22};
23
24/// Serve one session over stdin and stdout for the instance at `layout`.
25pub async fn run_stdio(layout: &Layout, overrides: ConfigOverrides) -> Result<()> {
26    let stdin = tokio::io::stdin();
27    let stdout = tokio::io::stdout();
28    let tasks = TaskTracker::new();
29    let registry = instance_delegations(layout);
30    // Without a daemon, a later `scv exec` is what cleans up after an earlier
31    // one that was killed; this runs alongside the session.
32    tokio::spawn(reconcile_delegations(Arc::clone(&registry)));
33    let result = run_managed(
34        stdin,
35        stdout,
36        Instance {
37            layout: layout.clone(),
38            overrides,
39        },
40        Shared {
41            components: None,
42            registry,
43            hosts: None,
44        },
45        CancellationToken::new(),
46        tasks.clone(),
47    )
48    .await;
49    tasks.close();
50    tasks.wait().await;
51    result
52}
53
54/// Run the authoritative server on the instance's Unix socket.
55pub async fn run_socket(layout: &Layout, overrides: ConfigOverrides) -> Result<()> {
56    let socket = layout.socket();
57    let path = socket.as_path();
58    let instance = Instance {
59        layout: layout.clone(),
60        overrides,
61    };
62    if let Some(parent) = path.parent() {
63        tokio::fs::create_dir_all(parent)
64            .await
65            .context("create SCV socket directory")?;
66        #[cfg(unix)]
67        {
68            use std::os::unix::fs::PermissionsExt;
69            std::fs::set_permissions(parent, std::fs::Permissions::from_mode(0o700))
70                .context("secure SCV socket directory")?;
71        }
72    }
73    let _lock = SocketLock::acquire(path)?;
74    for warning in instance.claude_warnings() {
75        tracing::warn!("{warning}");
76    }
77    // Nothing reads an older release's files; say so once rather than let
78    // them look like live configuration.
79    if let Ok(strays) = layout.strays() {
80        for stray in strays.into_iter().filter(|stray| stray.legacy) {
81            tracing::warn!(
82                "{} is from an older SCV layout and is not used; see `scv config show`",
83                stray.path.display()
84            );
85        }
86    }
87    if path.exists() {
88        if UnixStream::connect(path).await.is_ok() {
89            return Err(anyhow!(
90                "SCV server is already running at {}",
91                path.display()
92            ));
93        }
94        use std::os::unix::fs::FileTypeExt;
95        if !std::fs::symlink_metadata(path)?.file_type().is_socket() {
96            return Err(anyhow!(
97                "refusing to remove a non-socket at SCV socket path"
98            ));
99        }
100        tokio::fs::remove_file(path)
101            .await
102            .with_context(|| format!("remove stale SCV socket {}", path.display()))?;
103    }
104    let listener = UnixListener::bind(path)
105        .with_context(|| format!("bind SCV server socket {}", path.display()))?;
106    #[cfg(unix)]
107    {
108        use std::os::unix::fs::PermissionsExt;
109        std::fs::set_permissions(path, std::fs::Permissions::from_mode(0o600))
110            .context("secure SCV socket")?;
111    }
112    let hub = scv_channels::hub::Hub::new(Some(layout.last_owner()));
113    // Before any account starts: its recovery needs to know whether this
114    // start is a planned restart.
115    let startup = restart::startup(layout, &hub);
116    let components = Arc::new(Mutex::new(components::Components::with_hub(
117        instance.clone(),
118        std::env::current_dir()?,
119        Arc::clone(&hub),
120    )?));
121    let registry = instance_delegations(layout);
122    let hosts = crate::workbench::Hosts::new(instance.clone(), Arc::clone(&registry));
123    // Descendants a delegated agent leaves behind reparent to the daemon, not init.
124    if !delegations::become_child_subreaper() {
125        tracing::debug!("SCV daemon is not a child subreaper on this platform");
126    }
127    let cancellation = CancellationToken::new();
128    let restarter = restart::Restarter::new(
129        instance.clone(),
130        Arc::clone(&hub),
131        Arc::clone(&registry),
132        &components,
133        cancellation.clone(),
134    );
135    let disk = tokio::spawn(disk::monitor(
136        instance.clone(),
137        Arc::clone(&hub),
138        restarter.notifier().clone(),
139        cancellation.clone(),
140    ));
141    let _disk_abort = AbortGuard(disk.abort_handle());
142    let confirmer = confirm::Confirmer::new(
143        hub,
144        Arc::clone(&registry),
145        restarter.notifier().clone(),
146        cancellation.clone(),
147    );
148    {
149        let mut components = components.lock().await;
150        components.set_restarter(Arc::clone(&restarter));
151        components.set_confirmer(confirmer);
152    }
153    let notices = tokio::spawn(restart::announce(
154        layout.clone(),
155        startup,
156        restarter.notifier().clone(),
157        cancellation.clone(),
158    ));
159    let _notices_abort = AbortGuard(notices.abort_handle());
160    let monitor = tokio::spawn(restart::monitor(
161        restarter.notifier().clone(),
162        cancellation.clone(),
163    ));
164    let _monitor_abort = AbortGuard(monitor.abort_handle());
165    let delegation_registry = Arc::clone(&registry);
166    let delegation_cancel = cancellation.clone();
167    let maintained = Arc::clone(&hosts);
168    let mut delegation_task = tokio::spawn(async move {
169        // The first tick is immediate: orphans from before a restart go
170        // first, and jobs and runs they left unfinished become interrupted.
171        let mut interval = tokio::time::interval(DELEGATION_RECONCILE_INTERVAL);
172        loop {
173            tokio::select! {
174                biased;
175                () = delegation_cancel.cancelled() => break,
176                _ = interval.tick() => {
177                    reconcile_delegations(Arc::clone(&delegation_registry)).await;
178                    let zombies = delegations::reap_orphaned_zombies();
179                    if zombies > 0 {
180                        tracing::debug!("Reaped {zombies} exited orphan processes");
181                    }
182                    crate::workbench::maintain(&maintained).await;
183                }
184            }
185        }
186    });
187    let _delegation_abort = AbortGuard(delegation_task.abort_handle());
188    let tasks = TaskTracker::new();
189    let mut clients = tokio::task::JoinSet::new();
190    let refresh_components = components.clone();
191    let refresh_cancel = cancellation.clone();
192    let mut refresh_task = tokio::spawn(async move {
193        let mut refresh = tokio::time::interval(Duration::from_secs(2));
194        loop {
195            tokio::select! {
196                biased;
197                () = refresh_cancel.cancelled() => break,
198                _ = refresh.tick() => {
199                    tokio::select! {
200                        biased;
201                        () = refresh_cancel.cancelled() => break,
202                        result = async { refresh_components.lock().await.reconcile().await } => {
203                            if result.is_err() { tracing::warn!("Component account discovery failed"); }
204                        }
205                    }
206                }
207            }
208        }
209    });
210    let _refresh_abort = AbortGuard(refresh_task.abort_handle());
211    let mut terminate = tokio::signal::unix::signal(tokio::signal::unix::SignalKind::terminate())?;
212    let result = loop {
213        tokio::select! {
214            accepted = listener.accept() => {
215                let (stream, _) = match accepted { Ok(value) => value, Err(error) => break Err(error.into()) };
216                let instance = instance.clone();
217                let components = components.clone();
218                let registry = Arc::clone(&registry);
219                let hosts = Arc::clone(&hosts);
220                let cancellation = cancellation.clone();
221                let tasks = tasks.clone();
222                clients.spawn(async move {
223                    let (reader, writer) = stream.into_split();
224                    if let Err(error) = run_managed(reader, writer, instance, Shared { components: Some(components), registry, hosts: Some(hosts) }, cancellation, tasks).await {
225                        tracing::warn!(error = format!("{error:#}"), "SCV socket client stopped");
226                    }
227                });
228            }
229            _ = clients.join_next(), if !clients.is_empty() => {},
230            _ = tokio::signal::ctrl_c() => break Ok(()),
231            _ = terminate.recv() => break Ok(()),
232        }
233    };
234    drop(listener);
235    cancellation.cancel();
236    let _ = (&mut refresh_task).await;
237    let _ = (&mut delegation_task).await;
238    components.lock().await.shutdown().await;
239    // The owner's own jobs stop with the daemon, like every session's.
240    hosts.shutdown().await;
241    if tokio::time::timeout(Duration::from_secs(8), async {
242        while clients.join_next().await.is_some() {}
243    })
244    .await
245    .is_err()
246    {
247        clients.abort_all();
248        while clients.join_next().await.is_some() {}
249    }
250    tasks.close();
251    tasks.wait().await;
252    let _ = tokio::fs::remove_file(path).await;
253    restart::clean_shutdown(layout);
254    result
255}
256
257pub(crate) const DELEGATION_RECONCILE_INTERVAL: Duration = Duration::from_secs(60);
258
259/// The delegation registry for this process's SCV instance.
260pub(crate) fn instance_delegations(layout: &Layout) -> Arc<DelegationRegistry> {
261    Arc::new(DelegationRegistry::new(layout))
262}
263
264/// Stop orphaned delegations of this instance and log what was stopped.
265pub(crate) async fn reconcile_delegations(registry: Arc<DelegationRegistry>) {
266    let report = registry.reconcile().await;
267    if !report.reaped.is_empty() {
268        tracing::info!(
269            "Reaped {} orphaned delegations: {}",
270            report.reaped.len(),
271            report.reaped.join(", ")
272        );
273    }
274    if report.removed > 0 {
275        tracing::debug!(
276            "Removed {} delegation records whose processes had exited",
277            report.removed
278        );
279    }
280    if report.stale_markers > 0 {
281        tracing::debug!(
282            "Removed {} conversation markers whose SCV process had exited",
283            report.stale_markers
284        );
285    }
286}
287
288/// A persistent advisory lock closes the stale-socket unlink/bind race.
289pub(crate) struct SocketLock(std::fs::File);
290impl SocketLock {
291    pub(crate) fn acquire(socket: &Path) -> Result<Self> {
292        use std::os::unix::{fs::OpenOptionsExt, io::AsRawFd};
293        let file = std::fs::OpenOptions::new()
294            .read(true)
295            .write(true)
296            .create(true)
297            .truncate(false)
298            .mode(0o600)
299            .custom_flags(libc::O_NOFOLLOW)
300            .open(socket.with_extension("lock"))?;
301        // SAFETY: flock operates on this owned, live file descriptor.
302        if unsafe { libc::flock(file.as_raw_fd(), libc::LOCK_EX | libc::LOCK_NB) } != 0 {
303            return Err(anyhow!("SCV daemon already owns this socket"));
304        }
305        Ok(Self(file))
306    }
307}
308impl Drop for SocketLock {
309    fn drop(&mut self) {
310        use std::os::unix::io::AsRawFd;
311        // SAFETY: the descriptor remains live until this drop returns.
312        unsafe {
313            libc::flock(self.0.as_raw_fd(), libc::LOCK_UN);
314        }
315    }
316}
317
318/// Aborts a task when dropped, so a task the daemon spawned never outlives it.
319pub(crate) struct AbortGuard(pub(crate) tokio::task::AbortHandle);
320impl Drop for AbortGuard {
321    fn drop(&mut self) {
322        self.0.abort();
323    }
324}
325
326#[cfg(test)]
327mod tests;
scv_server/daemon.rs

scv_server/
daemon.rs
