kill notebook its self when server cull idle kernel

[levinxo]

As we know, kernelManager will periodic check if kernel is inactivity. When set MappingKernelManager.cull_connected=True, it will cull idle kernel with connections. Then, when NotebookApp.shutdown_no_activity_timeout has exhausted, jupyter will call tornado's ioloop.stop() to stop the whole server and release resources. Now we can find jupyter process hanging and will not exit automatic.

Detail info (debug with tcpdump / wireshark / strace / telnet):

When kernel killed by kernel manager, it reply client a msg with msg_type=shutdown_reply, but client does not handle this msg. Tornado server will wait for websocket connection produced by client, and never exit for a very long time.

When we telnet tornado's port 8888, its status is listening but not accept new connection.

Debug jupyter process with strace we can find it block in a event: poll([{fd=280, events=POLLIN}], 1, -1.

[Zsailer]

The changes here look good to me—but the tests are failing. @levinxo, can you verify that the tests pass locally?

[levinxo]

I have tested it locally by running python3 -m notebook.jstest services. it has the same error with travis-ci produced. I will work on it to find out why and try to fix it :)

[kevin-bates]

I can't really comment on the code change or test issue, but, from a functional standpoint, I've confirmed that these changes address the issue (after reproducing the issue prior to running with these changes).

Thanks @levinxo.