http: Fix threading issue causing duplicate WS messages

The problem presents itself when a JSON-RPC call triggers some event in core.
When this happens we have a thread outside of Tornado call `write_message`
interleaved with a potentially ongoing write if the JSON-RPC response.

To avoid this we now follow Tornado best practices and add callbacks to the
IOLoop to ensure that there isn't any interleaved access of Tornado state.

docs/changelog.rst

@@ -23,6 +23,8 @@ Bug fix release.
   where the scanner hangs on non-audio files like video. The scanner will now
   also let us know if we found any decodeable audio. (Fixes: :issue:`726`)
 
+- HTTP: Fix threading bug that would cause duplicate delivery of WS messages.
+
 
 v1.0.0 (2015-03-25)
 ===================

mopidy/http/handlers.py

@@ -5,6 +5,7 @@ import os
 import socket
 
 import tornado.escape
+import tornado.ioloop
 import tornado.web
 import tornado.websocket
 
@@ -65,6 +66,19 @@ def make_jsonrpc_wrapper(core_actor):
     )
 
 
+def _send_broadcast(client, msg):
+    # We could check for client.ws_connection, but we don't really
+    # care why the broadcast failed, we just want the rest of them
+    # to succeed, so catch everything.
+    try:
+        client.write_message(msg)
+    except Exception as e:
+        error_msg = encoding.locale_decode(e)
+        logger.debug('Broadcast of WebSocket message to %s failed: %s',
+                     client.request.remote_ip, error_msg)
+        # TODO: should this do the same cleanup as the on_message code?
+
+
 class WebSocketHandler(tornado.websocket.WebSocketHandler):
 
     # XXX This set is shared by all WebSocketHandler objects. This isn't
@@ -74,17 +88,12 @@ class WebSocketHandler(tornado.websocket.WebSocketHandler):
 
     @classmethod
     def broadcast(cls, msg):
+        # This can be called from outside the Tornado ioloop, so we need to
+        # safely cross the thread boundary by adding a callback to the loop.
+        loop = tornado.ioloop.IOLoop.current()
         for client in cls.clients:
-            # We could check for client.ws_connection, but we don't really
-            # care why the broadcast failed, we just want the rest of them
-            # to succeed, so catch everything.
-            try:
-                client.write_message(msg)
-            except Exception as e:
-                error_msg = encoding.locale_decode(e)
-                logger.debug('Broadcast of WebSocket message to %s failed: %s',
-                             client.request.remote_ip, error_msg)
-                # TODO: should this do the same cleanup as the on_message code?
+            # One callback per client to keep time we hold up the loop short
+            loop.add_callback(_send_broadcast, client, msg)
 
     def initialize(self, core):
         self.jsonrpc = make_jsonrpc_wrapper(core)