on_finish() is not called when the client closed first.

class Webbench(tornado.web.RequestHandler):
count = 0
fini_count = 0;
closed_count = 0;

@tornado.web.asynchronous
def get(self, *args, **kwargs):
    Webbench.count += 1
    logging.debug("Webbench client count %d" % Webbench.count)

    self.write("hello world " + str(Webbench.count))
    // the finish() cause on_finish() being called.
    // self.finish()

def on_finish(self):
    Webbench.fini_count += 1
    logging.debug("Webbench finished %d, closed %d, total %d" 
                  % (Webbench.fini_count, Webbench.closed_count, Webbench.count))

def on_connection_close(self):
    // when client reset or closed connection, the on_finish() was not called.
    tornado.web.RequestHandler.on_connection_close(self)

    Webbench.closed_count += 1
    logging.debug("Webbench closed %d, finished %d, total %d" 
                  % (Webbench.closed_count, Webbench.fini_count, Webbench.count))

Start with tornado.web.RequestHandler.on_finish and on_connection_close, then trace the request cleanup path when the client closes first. Reproduce the Webbench example and add a regression test that makes the expected callback behavior explicit.