Reactive deadlock management in storage area networks

Abstract

Systems and methods in accordance with various embodiments can detect and alleviate potential or actual deadlock of a storage switch or storage area network when attempting to write data to a mirrored virtual target. In accordance with one embodiment, a timer is started when a storage switch routes a write command to the physical targets corresponding to a virtual target of the write command. If each physical target does not return a transfer ready resource within a predetermined timeout period, the switch determines that a potential or actual deadlock has occurred. An abort command is sent to each of the physical devices. The abort command can clear the command from the targets and also free any allocated transfer ready resources. After receiving an acceptance response from each physical target, the state of the write command at the switch can be cleared. The write command can then be re-issued to the physical devices. In one embodiment, a queue depth for the virtual target can be lowered after failing to receive transfer ready resources from each target. By decreasing the queue depth, the maximum number of commands the switch will process for a virtual target is decreased to decrease the probability of future deadlock. In one embodiment, the queue depth is increased after no deadlock is detected for a period of time.