From a2d10694ccfcd12dad8796fc86966ea10ca3fc01 Mon Sep 17 00:00:00 2001
From: Emil Lauridsen <mine809@gmail.com>
Date: Fri, 27 Dec 2019 11:31:25 +0100
Subject: Consistent, hopefully robust, shutdown/cancelation story for cargo
 check subprocess

---
 crates/ra_cargo_watch/src/lib.rs | 66 ++++++++++++++++++++++++++++++----------
 1 file changed, 50 insertions(+), 16 deletions(-)

diff --git a/crates/ra_cargo_watch/src/lib.rs b/crates/ra_cargo_watch/src/lib.rs
index c86386610..70afd7f8a 100644
--- a/crates/ra_cargo_watch/src/lib.rs
+++ b/crates/ra_cargo_watch/src/lib.rs
@@ -2,7 +2,7 @@
 //! another compatible command (f.x. clippy) in a background thread and provide
 //! LSP diagnostics based on the output of the command.
 use cargo_metadata::Message;
-use crossbeam_channel::{select, unbounded, Receiver, RecvError, Sender, TryRecvError};
+use crossbeam_channel::{select, unbounded, Receiver, RecvError, Sender};
 use lsp_types::{
     Diagnostic, Url, WorkDoneProgress, WorkDoneProgressBegin, WorkDoneProgressEnd,
     WorkDoneProgressReport,
@@ -191,7 +191,8 @@ impl CheckWatcherState {
                 self.last_update_req.take();
                 self.shared.write().clear(task_send);
 
-                self.watcher.cancel();
+                // By replacing the watcher, we drop the previous one which
+                // causes it to shut down automatically.
                 self.watcher = WatchThread::new(&self.options, &self.workspace_root);
             }
         }
@@ -277,9 +278,11 @@ impl CheckWatcherState {
 /// doesn't provide a way to read sub-process output without blocking, so we
 /// have to wrap sub-processes output handling in a thread and pass messages
 /// back over a channel.
+/// The correct way to dispose of the thread is to drop it, on which the
+/// sub-process will be killed, and the thread will be joined.
 struct WatchThread {
+    handle: Option<JoinHandle<()>>,
     message_recv: Receiver<CheckEvent>,
-    cancel_send: Sender<()>,
 }
 
 enum CheckEvent {
@@ -302,9 +305,8 @@ impl WatchThread {
         args.extend(options.args.iter().cloned());
 
         let (message_send, message_recv) = unbounded();
-        let (cancel_send, cancel_recv) = unbounded();
         let enabled = options.enable;
-        std::thread::spawn(move || {
+        let handle = std::thread::spawn(move || {
             if !enabled {
                 return;
             }
@@ -316,24 +318,56 @@ impl WatchThread {
                 .spawn()
                 .expect("couldn't launch cargo");
 
-            message_send.send(CheckEvent::Begin).unwrap();
+            // If we trigger an error here, we will do so in the loop instead,
+            // which will break out of the loop, and continue the shutdown
+            let _ = message_send.send(CheckEvent::Begin);
+
             for message in cargo_metadata::parse_messages(command.stdout.take().unwrap()) {
-                match cancel_recv.try_recv() {
-                    Ok(()) | Err(TryRecvError::Disconnected) => {
-                        command.kill().expect("couldn't kill command");
+                let message = match message {
+                    Ok(message) => message,
+                    Err(err) => {
+                        log::error!("Invalid json from cargo check, ignoring: {}", err);
+                        continue;
                     }
-                    Err(TryRecvError::Empty) => (),
-                }
+                };
 
-                message_send.send(CheckEvent::Msg(message.unwrap())).unwrap();
+                match message_send.send(CheckEvent::Msg(message)) {
+                    Ok(()) => {}
+                    Err(_err) => {
+                        // The send channel was closed, so we want to shutdown
+                        break;
+                    }
+                }
             }
-            message_send.send(CheckEvent::End).unwrap();
+
+            // We can ignore any error here, as we are already in the progress
+            // of shutting down.
+            let _ = message_send.send(CheckEvent::End);
+
+            // It is okay to ignore the result, as it only errors if the process is already dead
+            let _ = command.kill();
+
+            // Again, we don't care about the exit status so just ignore the result
+            let _ = command.wait();
         });
-        WatchThread { message_recv, cancel_send }
+        WatchThread { handle: Some(handle), message_recv }
     }
+}
 
-    fn cancel(&self) {
-        let _ = self.cancel_send.send(());
+impl std::ops::Drop for WatchThread {
+    fn drop(&mut self) {
+        if let Some(handle) = self.handle.take() {
+            // Replace our reciever with dummy one, so we can drop and close the
+            // one actually communicating with the thread
+            let recv = std::mem::replace(&mut self.message_recv, crossbeam_channel::never());
+
+            // Dropping the original reciever initiates thread sub-process shutdown
+            drop(recv);
+
+            // Join the thread, it should finish shortly. We don't really care
+            // whether it panicked, so it is safe to ignore the result
+            let _ = handle.join();
+        }
     }
 }
 
-- 
cgit v1.2.3