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    instance.log_policy();
78    // Nothing reads an older release's files; say so once rather than let
79    // them look like live configuration.
80    if let Ok(strays) = layout.strays() {
81        for stray in strays.into_iter().filter(|stray| stray.legacy) {
82            tracing::warn!(
83                "{} is from an older SCV layout and is not used; see `scv config show`",
84                stray.path.display()
85            );
86        }
87    }
88    if path.exists() {
89        if UnixStream::connect(path).await.is_ok() {
90            return Err(anyhow!(
91                "SCV server is already running at {}",
92                path.display()
93            ));
94        }
95        use std::os::unix::fs::FileTypeExt;
96        if !std::fs::symlink_metadata(path)?.file_type().is_socket() {
97            return Err(anyhow!(
98                "refusing to remove a non-socket at SCV socket path"
99            ));
100        }
101        tokio::fs::remove_file(path)
102            .await
103            .with_context(|| format!("remove stale SCV socket {}", path.display()))?;
104    }
105    let listener = UnixListener::bind(path)
106        .with_context(|| format!("bind SCV server socket {}", path.display()))?;
107    #[cfg(unix)]
108    {
109        use std::os::unix::fs::PermissionsExt;
110        std::fs::set_permissions(path, std::fs::Permissions::from_mode(0o600))
111            .context("secure SCV socket")?;
112    }
113    let hub = scv_channels::hub::Hub::new(Some(layout.last_owner()));
114    // Before any account starts: its recovery needs to know whether this
115    // start is a planned restart.
116    let startup = restart::startup(layout, &hub);
117    let components = Arc::new(Mutex::new(components::Components::with_hub(
118        instance.clone(),
119        std::env::current_dir()?,
120        Arc::clone(&hub),
121    )?));
122    let registry = instance_delegations(layout);
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    // Every daemon session asks the owner through it.
149    let hosts = crate::workbench::Hosts::new(
150        instance.clone(),
151        Arc::clone(&registry),
152        Some(Arc::clone(&confirmer)),
153    );
154    {
155        let mut components = components.lock().await;
156        components.set_restarter(Arc::clone(&restarter));
157        components.set_confirmer(confirmer);
158    }
159    let notices = tokio::spawn(restart::announce(
160        layout.clone(),
161        startup,
162        restarter.notifier().clone(),
163        cancellation.clone(),
164    ));
165    let _notices_abort = AbortGuard(notices.abort_handle());
166    let monitor = tokio::spawn(restart::monitor(
167        restarter.notifier().clone(),
168        cancellation.clone(),
169    ));
170    let _monitor_abort = AbortGuard(monitor.abort_handle());
171    let delegation_registry = Arc::clone(&registry);
172    let delegation_cancel = cancellation.clone();
173    let maintained = Arc::clone(&hosts);
174    let mut delegation_task = tokio::spawn(async move {
175        // The first tick is immediate: orphans from before a restart go
176        // first, and jobs and runs they left unfinished become interrupted.
177        let mut interval = tokio::time::interval(DELEGATION_RECONCILE_INTERVAL);
178        loop {
179            tokio::select! {
180                biased;
181                () = delegation_cancel.cancelled() => break,
182                _ = interval.tick() => {
183                    reconcile_delegations(Arc::clone(&delegation_registry)).await;
184                    let zombies = delegations::reap_orphaned_zombies();
185                    if zombies > 0 {
186                        tracing::debug!("Reaped {zombies} exited orphan processes");
187                    }
188                    crate::workbench::maintain(&maintained).await;
189                }
190            }
191        }
192    });
193    let _delegation_abort = AbortGuard(delegation_task.abort_handle());
194    let tasks = TaskTracker::new();
195    let mut clients = tokio::task::JoinSet::new();
196    let refresh_components = components.clone();
197    let refresh_cancel = cancellation.clone();
198    let mut refresh_task = tokio::spawn(async move {
199        let mut refresh = tokio::time::interval(Duration::from_secs(2));
200        loop {
201            tokio::select! {
202                biased;
203                () = refresh_cancel.cancelled() => break,
204                _ = refresh.tick() => {
205                    tokio::select! {
206                        biased;
207                        () = refresh_cancel.cancelled() => break,
208                        result = async { refresh_components.lock().await.reconcile().await } => {
209                            if result.is_err() { tracing::warn!("Component account discovery failed"); }
210                        }
211                    }
212                }
213            }
214        }
215    });
216    let _refresh_abort = AbortGuard(refresh_task.abort_handle());
217    let mut terminate = tokio::signal::unix::signal(tokio::signal::unix::SignalKind::terminate())?;
218    let result = loop {
219        tokio::select! {
220            accepted = listener.accept() => {
221                let (stream, _) = match accepted { Ok(value) => value, Err(error) => break Err(error.into()) };
222                let instance = instance.clone();
223                let components = components.clone();
224                let registry = Arc::clone(&registry);
225                let hosts = Arc::clone(&hosts);
226                let cancellation = cancellation.clone();
227                let tasks = tasks.clone();
228                clients.spawn(async move {
229                    let (reader, writer) = stream.into_split();
230                    if let Err(error) = run_managed(reader, writer, instance, Shared { components: Some(components), registry, hosts: Some(hosts) }, cancellation, tasks).await {
231                        tracing::warn!(error = format!("{error:#}"), "SCV socket client stopped");
232                    }
233                });
234            }
235            _ = clients.join_next(), if !clients.is_empty() => {},
236            _ = tokio::signal::ctrl_c() => break Ok(()),
237            _ = terminate.recv() => break Ok(()),
238        }
239    };
240    drop(listener);
241    cancellation.cancel();
242    let _ = (&mut refresh_task).await;
243    let _ = (&mut delegation_task).await;
244    components.lock().await.shutdown().await;
245    // The owner's own jobs stop with the daemon, like every session's.
246    hosts.shutdown().await;
247    if tokio::time::timeout(Duration::from_secs(8), async {
248        while clients.join_next().await.is_some() {}
249    })
250    .await
251    .is_err()
252    {
253        clients.abort_all();
254        while clients.join_next().await.is_some() {}
255    }
256    tasks.close();
257    tasks.wait().await;
258    let _ = tokio::fs::remove_file(path).await;
259    restart::clean_shutdown(layout);
260    result
261}
262
263pub(crate) const DELEGATION_RECONCILE_INTERVAL: Duration = Duration::from_secs(60);
264
265/// The delegation registry for this process's SCV instance.
266pub(crate) fn instance_delegations(layout: &Layout) -> Arc<DelegationRegistry> {
267    Arc::new(DelegationRegistry::new(layout))
268}
269
270/// Stop orphaned delegations of this instance and log what was stopped.
271pub(crate) async fn reconcile_delegations(registry: Arc<DelegationRegistry>) {
272    let report = registry.reconcile().await;
273    if !report.reaped.is_empty() {
274        tracing::info!(
275            "Reaped {} orphaned delegations: {}",
276            report.reaped.len(),
277            report.reaped.join(", ")
278        );
279    }
280    if report.removed > 0 {
281        tracing::debug!(
282            "Removed {} delegation records whose processes had exited",
283            report.removed
284        );
285    }
286    if report.stale_markers > 0 {
287        tracing::debug!(
288            "Removed {} conversation markers whose SCV process had exited",
289            report.stale_markers
290        );
291    }
292}
293
294/// A persistent advisory lock closes the stale-socket unlink/bind race.
295pub(crate) struct SocketLock(std::fs::File);
296impl SocketLock {
297    pub(crate) fn acquire(socket: &Path) -> Result<Self> {
298        use std::os::unix::{fs::OpenOptionsExt, io::AsRawFd};
299        let file = std::fs::OpenOptions::new()
300            .read(true)
301            .write(true)
302            .create(true)
303            .truncate(false)
304            .mode(0o600)
305            .custom_flags(libc::O_NOFOLLOW)
306            .open(socket.with_extension("lock"))?;
307        // SAFETY: flock operates on this owned, live file descriptor.
308        if unsafe { libc::flock(file.as_raw_fd(), libc::LOCK_EX | libc::LOCK_NB) } != 0 {
309            return Err(anyhow!("SCV daemon already owns this socket"));
310        }
311        Ok(Self(file))
312    }
313}
314impl Drop for SocketLock {
315    fn drop(&mut self) {
316        use std::os::unix::io::AsRawFd;
317        // SAFETY: the descriptor remains live until this drop returns.
318        unsafe {
319            libc::flock(self.0.as_raw_fd(), libc::LOCK_UN);
320        }
321    }
322}
323
324/// Aborts a task when dropped, so a task the daemon spawned never outlives it.
325pub(crate) struct AbortGuard(pub(crate) tokio::task::AbortHandle);
326impl Drop for AbortGuard {
327    fn drop(&mut self) {
328        self.0.abort();
329    }
330}
331
332#[cfg(test)]
333mod tests;
scv_server/daemon.rs

scv_server/
daemon.rs
