Possible race condition on Windows platform causing task to never wake

[Andrepuel]

Context:

if-watch/src/win.rs

Line 99 in 5ad5fda

if !self.resync.swap(false, Ordering::Relaxed) {

The potential scenario with two threads:

1. Thread A runs line 99, resync AtomicBool returns false

   if-watch/src/win.rs

   Line 99 in 5ad5fda

   if !self.resync.swap(false, Ordering::Relaxed) {

2. Thread B, the win32 notifier, runs line 61, sets resync to true

   if-watch/src/win.rs

   Line 61 in 5ad5fda

   resync.store(true, Ordering::Relaxed);

3. Thread B, runs line 62, there is no watcher on the AtomicWaker yet

   if-watch/src/win.rs

   Line 62 in 5ad5fda

   waker.wake();

4. Thread A run lines 100, stores the Waker

   if-watch/src/win.rs

   Line 100 in 5ad5fda

   self.waker.register(cx.waker());

5. Thread A returns Pending

   if-watch/src/win.rs

   Line 101 in 5ad5fda

   return Poll::Pending;

The task running on Thread A will be pending forever, because resync is true while the call to waker.wake() was done before the Waker itself was registered.

Registering the Waker before loading resync would fix this issue, with the added side of effect of possible spurious wakes. However, spurious wakes should always be expected when dealing with Future. Another approach, would be to have both resync and waker inside a mutex, then we can read resync and set the waker on the same critical area.