We've had sporadic issues where a EBADF error when calling accept() was being logged. See example here:
https://travis-ci.org/dask/distributed/jobs/238031120#L1596-L1607

After some debugging, I've come to the conclusion that this happens whenever a IO handler calls TCPServer.stop(), while another handler is pending for a connection request on that TCPServer's listening socket(s). TCPServer.stop() closes the listening socket immediately and, when the queued accept handler is called, it finds a closed socket fd.