poll: Do not clear readiness on short read/writes. (#5881)

The new mio_unsupported_force_poll_poll behaviour works the same as
Windows (using level-triggered APIs to mimic edge-triggered ones) and it
depends on intercepting an EAGAIN result to start polling the fd again.

Author: jasta

tokio/src/io/poll_evented.rs

@@ -165,8 +165,10 @@ feature! {
                 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) {
+                        // that the socket buffer has been drained.  Unfortunately this assumption
+                        // fails for level-triggered selectors (like on Windows or poll even for
+                        // UNIX): https://github.com/tokio-rs/tokio/issues/5866
+                        if n > 0 && (!cfg!(windows) && !cfg!(mio_unsupported_force_poll_poll) && n < len) {
                             self.registration.clear_readiness(evt);
                         }
 
@@ -196,8 +198,10 @@ feature! {
                 match self.io.as_ref().unwrap().write(buf) {
                     Ok(n) => {
                         // if we write only part of our buffer, this is sufficient on unix to show
-                        // that the socket buffer is full
-                        if n > 0 && (!cfg!(windows) && n < buf.len()) {
+                        // that the socket buffer is full.  Unfortunately this assumption
+                        // fails for level-triggered selectors (like on Windows or poll even for
+                        // UNIX): https://github.com/tokio-rs/tokio/issues/5866
+                        if n > 0 && (!cfg!(windows) && !cfg!(mio_unsupported_force_poll_poll) && n < buf.len()) {
                             self.registration.clear_readiness(evt);
                         }