Prevent race condition in Notifier#run / #stop

There is a potential race condition when a notifier is run in a new
thread and then immediately stopped in the original thread.  If
`Notifier#stop` sets `@stop = true` before `Notifier#run` sets
`@stop = false`, then `Notifier#run` will loop until `Notifier#stop` is
called again.  Furthermore, if `Notifier#run` acquires the `@running`
mutex before `Notifier#stop` does (but after `Notifier#stop` sets
`@stop = true`), then `Notifier#stop` will get stuck waiting for the
mutex while `Notifier#run` infinitely loops.

This commit modifies `Notifier#run` to assume `@stop == false` when it
begins, and to reset `@stop = false` only after its loop terminates,
thus eliminating the race.

lib/rb-inotify/notifier.rb

@@ -50,6 +50,7 @@ def fd
     # @raise [SystemCallError] if inotify failed to initialize for some reason
     def initialize
       @running = Mutex.new
+      @stop = false
       @pipe = IO.pipe
       # JRuby shutdown sometimes runs IO finalizers before all threads finish.
       if RUBY_ENGINE == 'jruby'
@@ -231,12 +232,12 @@ def watch(path, *flags, &callback)
     def run
       @running.synchronize do
         Thread.current[:INOTIFY_RUN_THREAD] = true
-        @stop = false
 
         process until @stop
       end
     ensure
       Thread.current[:INOTIFY_RUN_THREAD] = false
+      @stop = false
     end
 
     # Stop watching for filesystem events.

spec/notifier_spec.rb

@@ -107,6 +107,20 @@
         dir.join("one.txt").write("hello world")
         run_thread.join(1) or raise "timeout"
       end
+
+      it "can be stopped before it starts processing" do
+        barrier = Concurrent::Event.new
+
+        run_thread = Thread.new do
+          barrier.wait
+          @notifier.run
+        end
+
+        @notifier.stop
+        barrier.set
+
+        run_thread.join(1) or raise "timeout"
+      end
     end
 
     describe :fd do