xous: std: thread_parking: fix deadlocks

Fix a deadlock condition that can occur when a thread is awoken in
between the point at which it checks its wake state and the point where
it actually waits.

This change will cause the waker to continuously send Notify messages
until it actually wakes up the target thread.

Signed-off-by: Sean Cross <sean@xobs.io>

library/std/src/sys/pal/xous/thread_parking.rs

@@ -29,31 +29,40 @@ impl Parker {
         // Change NOTIFIED to EMPTY and EMPTY to PARKED.
         let state = self.state.fetch_sub(1, Acquire);
         if state == NOTIFIED {
+            // The state has gone from NOTIFIED (1) to EMPTY (0)
             return;
         }
+        // The state has gone from EMPTY (0) to PARKED (-1)
+        assert!(state == EMPTY);
 
-        // The state was set to PARKED. Wait until the `unpark` wakes us up.
+        // The state is now PARKED (-1). Wait until the `unpark` wakes us up.
         blocking_scalar(
             ticktimer_server(),
             TicktimerScalar::WaitForCondition(self.index(), 0).into(),
         )
         .expect("failed to send WaitForCondition command");
 
-        self.state.swap(EMPTY, Acquire);
+        let state = self.state.swap(EMPTY, Acquire);
+        assert!(state == NOTIFIED || state == PARKED);
     }
 
     pub unsafe fn park_timeout(self: Pin<&Self>, timeout: Duration) {
         // Change NOTIFIED to EMPTY and EMPTY to PARKED.
         let state = self.state.fetch_sub(1, Acquire);
         if state == NOTIFIED {
+            // The state has gone from NOTIFIED (1) to EMPTY (0)
             return;
         }
+        // The state has gone from EMPTY (0) to PARKED (-1)
+        assert!(state == EMPTY);
 
         // A value of zero indicates an indefinite wait. Clamp the number of
         // milliseconds to the allowed range.
         let millis = usize::max(timeout.as_millis().try_into().unwrap_or(usize::MAX), 1);
 
-        let was_timeout = blocking_scalar(
+        // The state is now PARKED (-1). Wait until the `unpark` wakes us up,
+        // or things time out.
+        let _was_timeout = blocking_scalar(
             ticktimer_server(),
             TicktimerScalar::WaitForCondition(self.index(), millis).into(),
         )
@@ -61,28 +70,37 @@ impl Parker {
             != 0;
 
         let state = self.state.swap(EMPTY, Acquire);
-        if was_timeout && state == NOTIFIED {
-            // The state was set to NOTIFIED after we returned from the wait
-            // but before we reset the state. Therefore, a wakeup is on its
-            // way, which we need to consume here.
-            // NOTICE: this is a priority hole.
-            blocking_scalar(
-                ticktimer_server(),
-                TicktimerScalar::WaitForCondition(self.index(), 0).into(),
-            )
-            .expect("failed to send WaitForCondition command");
-        }
+        assert!(state == PARKED || state == NOTIFIED);
     }
 
     pub fn unpark(self: Pin<&Self>) {
-        let state = self.state.swap(NOTIFIED, Release);
-        if state == PARKED {
-            // The thread is parked, wake it up.
-            blocking_scalar(
-                ticktimer_server(),
-                TicktimerScalar::NotifyCondition(self.index(), 1).into(),
-            )
-            .expect("failed to send NotifyCondition command");
+        // If the state is already `NOTIFIED`, then another thread has
+        // indicated it wants to wake up the target thread.
+        //
+        // If the state is `EMPTY` then there is nothing to wake up, and
+        // the target thread will immediately exit from `park()` the
+        // next time that function is called.
+        if self.state.swap(NOTIFIED, Release) != PARKED {
+            return;
+        }
+
+        // The thread is parked, wake it up. Keep trying until we wake something up.
+        // This will happen when the `NotifyCondition` call returns the fact that
+        // 1 condition was notified.
+        // Alternately, keep going until the state is seen as `EMPTY`, indicating
+        // the thread woke up and kept going. This can happen when the Park
+        // times out before we can send the NotifyCondition message.
+        while blocking_scalar(
+            ticktimer_server(),
+            TicktimerScalar::NotifyCondition(self.index(), 1).into(),
+        )
+        .expect("failed to send NotifyCondition command")[0]
+            == 0
+            && self.state.load(Acquire) != EMPTY
+        {
+            // The target thread hasn't yet hit the `WaitForCondition` call.
+            // Yield to let the target thread run some more.
+            crate::thread::yield_now();
         }
     }
 }