See SERVER-55486 for more details.

The idea is that when the disk fails, the primary can spend long time without stepping down. At some point, presumably, it is moved to uninterruptible sleep state. It may or may not be related to the bug, granted. After being killed with SIGKILL the election happens, but mongos remains stuck with the old primary and never recovers unless killed.

The behavior reproduced in 4.0, but could be present in other branches.