perf(driver): set setup_coop_taskrun for io-uring (#401)

* perf(driver): set setup_coop_taskrun for io-uring

this can optimize performance for most cases

* feat(driver): allow user to choose enable coop_taskrun and taskrun_flag or not

Author: Sherlock-Holo

compio-driver/src/iour/mod.rs

@@ -100,6 +100,13 @@ impl Driver {
         if let Some(sqpoll_idle) = builder.sqpoll_idle {
             io_uring_builder.setup_sqpoll(sqpoll_idle.as_millis() as _);
         }
+        if builder.coop_taskrun {
+            io_uring_builder.setup_coop_taskrun();
+        }
+        if builder.taskrun_flag {
+            io_uring_builder.setup_taskrun_flag();
+        }
+
         let mut inner = io_uring_builder.build(builder.capacity)?;
         #[allow(clippy::useless_conversion)]
         unsafe {
@@ -127,14 +134,23 @@ impl Driver {
     // Auto means that it choose to wait or not automatically.
     fn submit_auto(&mut self, timeout: Option<Duration>) -> io::Result<()> {
         instrument!(compio_log::Level::TRACE, "submit_auto", ?timeout);
+
+        // when taskrun is true, there are completed cqes wait to handle, no need to
+        // block the submit
+        let want_sqe = if self.inner.submission().taskrun() {
+            0
+        } else {
+            1
+        };
+
         let res = {
             // Last part of submission queue, wait till timeout.
             if let Some(duration) = timeout {
                 let timespec = timespec(duration);
                 let args = SubmitArgs::new().timespec(&timespec);
-                self.inner.submitter().submit_with_args(1, &args)
+                self.inner.submitter().submit_with_args(want_sqe, &args)
             } else {
-                self.inner.submit_and_wait(1)
+                self.inner.submit_and_wait(want_sqe)
             }
         };
         trace!("submit result: {res:?}");

compio-driver/src/lib.rs

@@ -438,6 +438,8 @@ pub struct ProactorBuilder {
     capacity: u32,
     pool_builder: ThreadPoolBuilder,
     sqpoll_idle: Option<Duration>,
+    coop_taskrun: bool,
+    taskrun_flag: bool,
 }
 
 impl Default for ProactorBuilder {
@@ -453,6 +455,8 @@ impl ProactorBuilder {
             capacity: 1024,
             pool_builder: ThreadPoolBuilder::new(),
             sqpoll_idle: None,
+            coop_taskrun: false,
+            taskrun_flag: false,
         }
     }
 
@@ -516,6 +520,33 @@ impl ProactorBuilder {
         self
     }
 
+    /// `coop_taskrun` feature has been available since Linux Kernel 5.19. This
+    /// will optimize performance for most cases, especially compio is a single
+    /// thread runtime.
+    ///
+    /// However, it can't run with sqpoll feature.
+    ///
+    /// # Notes
+    ///
+    /// - Only effective when the `io-uring` feature is enabled
+    pub fn coop_taskrun(&mut self, enable: bool) -> &mut Self {
+        self.coop_taskrun = enable;
+        self
+    }
+
+    /// `taskrun_flag` feature has been available since Linux Kernel 5.19. This
+    /// allows io-uring driver can know if any cqes are available when try to
+    /// push sqe to sq. This should be enabled with
+    /// [`coop_taskrun`](Self::coop_taskrun)
+    ///
+    /// # Notes
+    ///
+    /// - Only effective when the `io-uring` feature is enabled
+    pub fn taskrun_flag(&mut self, enable: bool) -> &mut Self {
+        self.taskrun_flag = enable;
+        self
+    }
+
     /// Build the [`Proactor`].
     pub fn build(&self) -> io::Result<Proactor> {
         Proactor::with_builder(self)