Fix potential race condition during disconnection (#2719)

When the disconnect() function is called twice in parallel it is possible that
one thread deletes the self._sock reference, while the other thread will
attempt to call .close() on it, leading to an AttributeError.

This situation can routinely be encountered by closing the connection in a
PubSubWorkerThread error handler in a blocking thread (ie. with
sleep_time==None), and then calling .close() on the PubSub object.
The main thread will then run into the disconnect() function, and the listener
thread is woken up by the closure and will race into the disconnect()
function, too.

This can be fixed easily by copying the object reference before doing the
None-check, similar to what we do in the redis.client.close() function.

redis/connection.py

@@ -779,20 +779,22 @@ def on_connect(self):
     def disconnect(self, *args):
         "Disconnects from the Redis server"
         self._parser.on_disconnect()
-        if self._sock is None:
+
+        conn_sock = self._sock
+        self._sock = None
+        if conn_sock is None:
             return
 
         if os.getpid() == self.pid:
             try:
-                self._sock.shutdown(socket.SHUT_RDWR)
+                conn_sock.shutdown(socket.SHUT_RDWR)
             except OSError:
                 pass
 
         try:
-            self._sock.close()
+            conn_sock.close()
         except OSError:
             pass
-        self._sock = None
 
     def _send_ping(self):
         """Send PING, expect PONG in return"""