Use lock-free storage for runtime thread queue

Cargo.lock

@@ -517,6 +517,7 @@ dependencies = [
 name = "smol-mlua"
 version = "0.0.0"
 dependencies = [
+ "concurrent-queue",
  "mlua",
  "smol",
 ]

Cargo.toml

@@ -4,6 +4,7 @@ version = "0.0.0"
 edition = "2021"
 
 [dependencies]
+concurrent-queue = "2.4"
 smol = "2.0"
 mlua = { version = "0.9", features = ["luau", "luau-jit", "async"] }
 

lib/queue.rs

@@ -1,13 +1,8 @@
-use std::sync::{
-    atomic::{AtomicBool, Ordering},
-    Arc,
-};
+use std::sync::Arc;
 
+use concurrent_queue::ConcurrentQueue;
 use mlua::prelude::*;
-use smol::{
-    channel::{unbounded, Receiver, Sender},
-    lock::Mutex,
-};
+use smol::channel::{unbounded, Receiver, Sender};
 
 use crate::IntoLuaThread;
 
@@ -21,27 +16,22 @@ const ERR_OOM: &str = "out of memory";
 */
 #[derive(Debug, Clone)]
 pub struct ThreadQueue {
-    queue: Arc<Mutex<Vec<ThreadWithArgs>>>,
-    status: Arc<AtomicBool>,
+    queue: Arc<ConcurrentQueue<ThreadWithArgs>>,
     signal_tx: Sender<()>,
     signal_rx: Receiver<()>,
 }
 
 impl ThreadQueue {
     pub fn new() -> Self {
+        let queue = Arc::new(ConcurrentQueue::unbounded());
         let (signal_tx, signal_rx) = unbounded();
         Self {
-            queue: Arc::new(Mutex::new(Vec::new())),
-            status: Arc::new(AtomicBool::new(false)),
+            queue,
             signal_tx,
             signal_rx,
         }
     }
 
-    pub fn has_threads(&self) -> bool {
-        self.status.load(Ordering::SeqCst)
-    }
-
     pub fn push<'lua>(
         &self,
         lua: &'lua Lua,
@@ -52,21 +42,23 @@ impl ThreadQueue {
         let args = args.into_lua_multi(lua)?;
         let stored = ThreadWithArgs::new(lua, thread, args);
 
-        self.queue.lock_blocking().push(stored);
-        self.status.store(true, Ordering::SeqCst);
+        self.queue.push(stored).unwrap();
         self.signal_tx.try_send(()).unwrap();
 
         Ok(())
     }
 
-    pub async fn drain<'lua>(&self, lua: &'lua Lua) -> Vec<(LuaThread<'lua>, LuaMultiValue<'lua>)> {
-        let mut queue = self.queue.lock().await;
-        let drained = queue.drain(..).map(|s| s.into_inner(lua)).collect();
-        self.status.store(false, Ordering::SeqCst);
-        drained
+    pub fn drain<'outer, 'lua>(
+        &'outer self,
+        lua: &'lua Lua,
+    ) -> impl Iterator<Item = (LuaThread<'lua>, LuaMultiValue<'lua>)> + 'outer
+    where
+        'lua: 'outer,
+    {
+        self.queue.try_iter().map(|stored| stored.into_inner(lua))
     }
 
-    pub async fn recv(&self) {
+    pub async fn listen(&self) {
         self.signal_rx.recv().await.unwrap();
         // Drain any pending receives
         loop {

lib/runtime.rs

@@ -184,8 +184,8 @@ impl<'lua> Runtime<'lua> {
             loop {
                 // Wait for a new thread to arrive __or__ next futures step, prioritizing
                 // new threads, so we don't accidentally exit when there is more work to do
-                let fut_spawn = self.queue_spawn.recv();
-                let fut_defer = self.queue_defer.recv();
+                let fut_spawn = self.queue_spawn.listen();
+                let fut_defer = self.queue_defer.listen();
                 let fut_tick = async {
                     lua_exec.tick().await;
                     // Do as much work as possible
@@ -200,11 +200,7 @@ impl<'lua> Runtime<'lua> {
 
                 // If a new thread was spawned onto any queue,
                 // we must drain them and schedule on the executor
-                if self.queue_spawn.has_threads() || self.queue_defer.has_threads() {
-                    let mut queued_threads = Vec::new();
-                    queued_threads.extend(self.queue_spawn.drain(self.lua).await);
-                    queued_threads.extend(self.queue_defer.drain(self.lua).await);
-                    for (thread, args) in queued_threads {
+                let process_thread = |thread: LuaThread<'lua>, args| {
                     // NOTE: Thread may have been cancelled from lua
                     // before we got here, so we need to check it again
                     if thread.status() == LuaThreadStatus::Resumable {
@@ -222,7 +218,14 @@ impl<'lua> Runtime<'lua> {
                             })
                             .detach();
                     }
+                };
+
+                // Process spawned threads first, then deferred threads
+                for (thread, args) in self.queue_spawn.drain(self.lua) {
+                    process_thread(thread, args);
                 }
+                for (thread, args) in self.queue_defer.drain(self.lua) {
+                    process_thread(thread, args);
                 }
 
                 // Empty executor = no remaining threads