We consider a wireless downlink network with a single base-station, N mobile users and L shared on-off channels. Each mobile user receives a downlink traffic flow from the basestation where a separate queue is maintained for each flow. In this multi-channel downlink network, throughput-optimal scheduling algorithms such as the MaxWeight scheduling require the complete channel state information (i.e., NL channel states) for scheduling. This could be a significant overhead when the number of mobile users is large. This paper considers wireless downlink networks with limited feedback bandwidth so that at most F of the NL channel states can be reported at each time slot. We propose dynamic feedback allocation schemes, named as Longest-Queue-First Feedback-Allocation (LQF-FA) and Modified-Longest-Queue-First Feedback-Allocation (MLQF-FA), which dynamically and adaptively allocate the feedback resource according to queue-lengths and channel statistics. We prove that given a fixed feedback resource F, the LQF-FA+MaxWeight is throughput-optimal under a mean approximation; and the throughput difference between the MLQF-FA+MaxWeight and the MaxWeight with the complete channel state information decreases exponentially as a function of F/L when F = O(L 2).