fix(host): don't block the event loop while a closing project shuts down (#1419)

Window destruction and the close_project command detach the project at once and let its language servers, kernels and debug sessions shut down in the background; opening a project waits for pending shutdowns and exit waits for them with a bound.

Closes #1414

crates/ide-host/src/lib.rs

@@ -84,8 +84,13 @@ pub struct Host {
     pub(crate) external_logs: RwLock<HashMap<String, Vec<Arc<ide_vcs::Repo>>>>,
     requests: requests::Requests,
     project_requests: RwLock<HashMap<String, Vec<RequestId>>>,
+    /// Shutdowns of projects closed with [`Host::close_project_detached`].
+    closing: parking_lot::Mutex<Vec<std::thread::JoinHandle<()>>>,
 }
 
+/// How long exit waits for project shutdown before giving up.
+const EXIT_SHUTDOWN_WAIT: std::time::Duration = std::time::Duration::from_secs(10);
+
 impl Host {
     pub fn new(config: HostConfig) -> Self {
         let HostConfig {
@@ -129,6 +134,7 @@ impl Host {
             external_logs: RwLock::default(),
             requests: Default::default(),
             project_requests: RwLock::default(),
+            closing: Default::default(),
         };
         host.apply_proxy();
         host
@@ -182,46 +188,69 @@ impl Host {
         }
     }
 
+    /// Closes `window`'s project and waits until its language servers, kernels and
+    /// debug sessions have shut down. Never call this on the UI event-loop thread:
+    /// the shutdown runs on the async runtime, whose workers may themselves be
+    /// waiting for the event loop (IPC responses, window queries). Use
+    /// [`Self::close_project_detached`] there.
     pub fn close_project(&self, window: &str) {
+        if let Some(shutdown) = self.begin_close(window) {
+            self.wait_blocking(shutdown);
+        }
+    }
+
+    /// Like [`Self::close_project`], but returns once the project is detached from
+    /// the window; its workers finish shutting down in the background, and
+    /// [`Self::shutdown_notebooks_on_exit`] waits for them before the process exits.
+    pub fn close_project_detached(&self, window: &str) {
+        if let Some(shutdown) = self.begin_close(window) {
+            let mut closing = self.closing.lock();
+            closing.retain(|handle| !handle.is_finished());
+            closing.push(shutdown);
+        }
+    }
+
+    /// Detaches the project from `window` and stops what it started, synchronously,
+    /// so the window can open another project at once. Returns the thread that
+    /// finishes shutting down its workers.
+    fn begin_close(&self, window: &str) -> Option<std::thread::JoinHandle<()>> {
         self.external_logs.write().remove(self.owner_window(window));
         for request in self.project_requests.write().remove(self.owner_window(window)).unwrap_or_default() {
             self.cancel_request(request);
         }
-        let project = self.projects.write().remove(window);
-        if let Some(project) = project {
-            if !project.light {
-                self.environment.project_closed();
-            }
-            if let Some(bridge) = &project.kotlin_bridge {
-                bridge.stop();
-            }
-            project.documents.clear_workspace_edit_history();
-            project.agents.shutdown();
-            project.notebooks.stop_admission();
-            // What the closing window started goes with it — the builds and searches it
-            // asked for, its runs, its shells and its debug sessions — and the language
-            // servers finish shutting down before the window is gone, as IDEA closes a
-            // project's processes.
-            self.processes.kill_project(&project.root);
-            self.terminals.kill_key(self.owner_window(window));
-            self.shutdown_project_workers_blocking(
-                vec![project.notebooks.clone()],
-                project.kotlin_bridge.iter().cloned().collect(),
-                Some(project.lsp.clone()),
-                Some((self.debugger.clone(), window.to_owned())),
-            );
+        let project = self.projects.write().remove(window)?;
+        if !project.light {
+            self.environment.project_closed();
         }
+        if let Some(bridge) = &project.kotlin_bridge {
+            bridge.stop();
+        }
+        project.documents.clear_workspace_edit_history();
+        project.agents.shutdown();
+        project.notebooks.stop_admission();
+        // What the closing window started goes with it — the builds and searches it
+        // asked for, its runs, its shells and its debug sessions — and its language
+        // servers shut down.
+        self.processes.kill_project(&project.root);
+        self.terminals.kill_key(self.owner_window(window));
+        Some(self.spawn_project_shutdown(
+            vec![project.notebooks.clone()],
+            project.kotlin_bridge.iter().cloned().collect(),
+            Some(project.lsp.clone()),
+            Some((self.debugger.clone(), window.to_owned())),
+        ))
     }
 
-    /// Synchronous project/window destruction must finish owned kernel cleanup,
-    /// not enqueue work which can be lost as the desktop runtime exits.
-    fn shutdown_project_workers_blocking(
+    /// Project/window destruction must finish owned kernel cleanup, not enqueue
+    /// work which can be lost as the desktop runtime exits: the returned thread is
+    /// joined by the closer or, for a detached close, before exit.
+    fn spawn_project_shutdown(
         &self,
         notebooks: Vec<Arc<notebook::NotebookService>>,
         bridges: Vec<Arc<kotlin_bridge::KotlinBridge>>,
         lsp: Option<Arc<ide_lsp::LspManager>>,
         debug: Option<(Arc<ide_dap::DebugManager>, String)>,
-    ) {
+    ) -> std::thread::JoinHandle<()> {
         for bridge in &bridges {
             bridge.stop();
         }
@@ -229,25 +258,37 @@ impl Host {
             service.stop_admission();
         }
         let runtime = self.runtime.clone();
-        let wait = move || {
-            let worker = std::thread::spawn(move || {
-                runtime.block_on(async move {
-                    for bridge in bridges {
-                        bridge.shutdown().await;
-                    }
-                    for service in notebooks {
-                        service.shutdown().await;
-                    }
-                    if let Some(lsp) = lsp {
-                        lsp.shutdown_all().await;
-                    }
-                    if let Some((debug, key)) = debug {
-                        debug.stop_key(&key).await;
-                    }
-                });
+        std::thread::spawn(move || {
+            runtime.block_on(async move {
+                for bridge in bridges {
+                    bridge.shutdown().await;
+                }
+                for service in notebooks {
+                    service.shutdown().await;
+                }
+                if let Some(lsp) = lsp {
+                    lsp.shutdown_all().await;
+                }
+                if let Some((debug, key)) = debug {
+                    debug.stop_key(&key).await;
+                }
             });
-            if let Err(error) = worker.join() {
-                tracing::error!("notebook shutdown thread failed: {error:?}");
+        })
+    }
+
+    /// Waits for detached closes, so a project opened next never races the
+    /// language servers of one still shutting down (e.g. over a shared workspace).
+    pub(crate) fn wait_detached_closes(&self) {
+        let pending: Vec<_> = self.closing.lock().drain(..).collect();
+        for shutdown in pending {
+            self.wait_blocking(shutdown);
+        }
+    }
+
+    fn wait_blocking(&self, shutdown: std::thread::JoinHandle<()>) {
+        let wait = move || {
+            if let Err(error) = shutdown.join() {
+                tracing::error!("project shutdown thread failed: {error:?}");
             }
         };
         if tokio::runtime::Handle::try_current()
@@ -259,15 +300,32 @@ impl Host {
         }
     }
 
-    /// Called before the desktop runtime exits, including update/restart exits.
+    /// Called before the desktop runtime exits, including update/restart exits, on
+    /// the event-loop thread: waits for detached closes and the open projects'
+    /// kernels, but no longer than [`EXIT_SHUTDOWN_WAIT`], so a runtime worker
+    /// stuck on the event loop cannot keep the process from exiting.
     pub fn shutdown_notebooks_on_exit(&self) {
         let projects = self.projects();
-        self.shutdown_project_workers_blocking(
+        let mut pending: Vec<_> = self.closing.lock().drain(..).collect();
+        pending.push(self.spawn_project_shutdown(
             projects.iter().map(|project| project.notebooks.clone()).collect(),
             projects.iter().filter_map(|project| project.kotlin_bridge.clone()).collect(),
             None,
             None,
-        );
+        ));
+        let deadline = std::time::Instant::now() + EXIT_SHUTDOWN_WAIT;
+        while pending.iter().any(|handle| !handle.is_finished()) {
+            if std::time::Instant::now() >= deadline {
+                tracing::warn!("project shutdown did not finish within {EXIT_SHUTDOWN_WAIT:?}; exiting anyway");
+                return;
+            }
+            std::thread::sleep(std::time::Duration::from_millis(20));
+        }
+        for handle in pending {
+            if let Err(error) = handle.join() {
+                tracing::error!("project shutdown thread failed: {error:?}");
+            }
+        }
     }
 
     /// Called before desktop exit, which may bypass document destructors.

crates/ide-host/src/project.rs

@@ -88,6 +88,7 @@ impl Host {
             return Ok(current);
         }
         self.close_project(window);
+        self.wait_detached_closes();
         let sink: SharedSink = (self.sink_for)(window);
         // Before anything the project starts: its children get the project's environment.
         let trusted = Arc::new(AtomicBool::new(self.trust.trusted(&root)?));
@@ -570,6 +571,7 @@ impl Host {
             return Ok(current);
         }
         self.close_project(window);
+        self.wait_detached_closes();
         let sink: SharedSink = (self.sink_for)(window);
         let settings = Arc::new(SettingsStore::load(&self.config_dir, None));
         let effective = settings.effective();

crates/ide-host/tests/close_project.rs

@@ -5,6 +5,7 @@
 use ide_core::NullSink;
 use ide_host::{Host, HostConfig, ops};
 use std::sync::Arc;
+use std::time::{Duration, Instant};
 
 fn options(root: &std::path::Path) -> ide_terminal::TerminalOptions {
     ide_terminal::TerminalOptions {
@@ -52,3 +53,53 @@ fn closing_a_project_closes_its_terminals_and_keeps_other_windows() {
         host.terminals.kill_key("other");
     });
 }
+
+/// The desktop closes a project on the event-loop thread while runtime workers may be
+/// blocked waiting for that thread (IPC responses, window queries): closing must not
+/// wait for workers that need the runtime, and exit still waits for them once free.
+#[test]
+fn detached_close_returns_while_runtime_workers_wait_for_the_closing_thread() {
+    let tmp = tempfile::tempdir().unwrap();
+    let root = dunce::canonicalize(tmp.path()).unwrap().join("project");
+    std::fs::create_dir_all(&root).unwrap();
+    let runtime = tokio::runtime::Builder::new_multi_thread().worker_threads(1).enable_all().build().unwrap();
+    let _guard = runtime.enter();
+    let host = Arc::new(Host::new(HostConfig {
+        config_dir: tmp.path().join("config"),
+        data_dir: tmp.path().join("data"),
+        local_data_dir: tmp.path().join("data"),
+        languages: Arc::new(ide_language::LanguageRegistry::load(None).unwrap()),
+        syntax: Arc::default(),
+        sink_for: Box::new(|_| Arc::new(NullSink)),
+        owner: |w| w,
+        runtime: runtime.handle().clone(),
+        project_environment: false,
+    }));
+    host.trust.decide(&root, true, false).unwrap();
+    host.open_project("main", &root).unwrap();
+
+    // Occupy the only runtime worker until "the event loop" answers.
+    let (started_tx, started_rx) = std::sync::mpsc::channel();
+    let (release_tx, release_rx) = std::sync::mpsc::channel::<()>();
+    runtime.spawn(async move {
+        started_tx.send(()).unwrap();
+        let _ = release_rx.recv();
+    });
+    started_rx.recv().unwrap();
+
+    let (closed_tx, closed_rx) = std::sync::mpsc::channel();
+    let closer = host.clone();
+    std::thread::spawn(move || {
+        closer.close_project_detached("main");
+        let _ = closed_tx.send(());
+    });
+    let returned = closed_rx.recv_timeout(Duration::from_secs(5));
+    release_tx.send(()).unwrap();
+    assert!(returned.is_ok(), "closing waited for a runtime worker that waits for the closing thread");
+    assert!(host.project("main").is_err(), "the window no longer has the project");
+    host.open_project("main", &root).expect("the window opens a project once the close has finished");
+
+    let started = Instant::now();
+    host.shutdown_notebooks_on_exit();
+    assert!(started.elapsed() < Duration::from_secs(5), "exit joined the finished close instead of timing out");
+}

src-tauri/src/state.rs

@@ -154,9 +154,11 @@ impl AppState {
         self.host.open_light(window, file)
     }
 
+    /// Called on the event-loop thread (window destruction, the `close_project`
+    /// command), so it must not wait for the project's workers to shut down.
     pub fn close_project(&self, window: &str) {
         self.remote_dev.disconnect(window);
-        self.host.close_project(window);
+        self.host.close_project_detached(window);
     }
 }