The growing reliance of networked applications on timely and reliable data transfer requires the underlying networking infrastructure to provide adequate services even in the presence of "persistent" failures (e.g., broken links/routers). It is much more difficult to meet this requirement for multicast sessions than for unicast communications because any on-tree component failure may often cause simultaneous service disruptions to multiple receivers. This paper presents a new multicast routing protocol, called the Survivable Multicast Routing Protocol (SMRP), which facilitates fast recovery of multicast sessions in face of persistent failures via local detour paths. Our evaluation results show that SMRP trades end-to-end delay and resource usage for short, and hence fast, recovery paths. For example, under a certain set of parameter values, SMRP shortens the recovery path by 20% with only a 5% increase of end-to-end delay and resource usage. Moreover, several design enhancements have made SMRP efficient, robust, flexible and scalable.