io: reduce syscalls in poll_read (#4840)

As the [epoll documentation points out](https://man7.org/linux/man-pages/man7/epoll.7.html), a read that only partially fills a buffer is sufficient to show that the socket buffer has been drained.

We can take advantage of this by clearing readiness in this case, which seems to significantly improve performance under certain conditions.

Co-authored-by: Carl Lerche <[email protected]>

tokio/src/io/driver/registration.rs

@@ -115,6 +115,7 @@ impl Registration {
 
     // Uses the poll path, requiring the caller to ensure mutual exclusion for
     // correctness. Only the last task to call this function is notified.
+    #[cfg(not(all(target_arch = "wasm32", target_os = "wasi")))]
     pub(crate) fn poll_read_io<R>(
         &self,
         cx: &mut Context<'_>,

tokio/src/io/poll_evented.rs

@@ -153,16 +153,32 @@ feature! {
         {
             use std::io::Read;
 
-            let n = ready!(self.registration.poll_read_io(cx, || {
-                let b = &mut *(buf.unfilled_mut() as *mut [std::mem::MaybeUninit<u8>] as *mut [u8]);
-                self.io.as_ref().unwrap().read(b)
-            }))?;
+            loop {
+                let evt = ready!(self.registration.poll_read_ready(cx))?;
 
-            // Safety: We trust `TcpStream::read` to have filled up `n` bytes in the
-            // buffer.
-            buf.assume_init(n);
-            buf.advance(n);
-            Poll::Ready(Ok(()))
+                let b = &mut *(buf.unfilled_mut() as *mut [std::mem::MaybeUninit<u8>] as *mut [u8]);
+                let len = b.len();
+
+                match self.io.as_ref().unwrap().read(b) {
+                    Ok(n) => {
+                        // if we read a partially full buffer, this is sufficient on unix to show
+                        // that the socket buffer has been drained
+                        if n > 0 && (!cfg!(windows) && n < len) {
+                            self.registration.clear_readiness(evt);
+                        }
+
+                        // Safety: We trust `TcpStream::read` to have filled up `n` bytes in the
+                        // buffer.
+                        buf.assume_init(n);
+                        buf.advance(n);
+                        return Poll::Ready(Ok(()));
+                    },
+                    Err(e) if e.kind() == io::ErrorKind::WouldBlock => {
+                        self.registration.clear_readiness(evt);
+                    }
+                    Err(e) => return Poll::Ready(Err(e)),
+                }
+            }
         }
 
         pub(crate) fn poll_write<'a>(&'a self, cx: &mut Context<'_>, buf: &[u8]) -> Poll<io::Result<usize>>