Epoptes clients delay 1.5 minutes on reboot/shutdown

[fottsia]

If I boot an LTSP client and immediately try to reboot it with epoptes, it works fine.

But if I do "lock screen" or "send a message", before logging in on the client, and any time after that I try reboot/shutdown, it delays 1.5 minute before doing so.

[alkisg]

The root epoptes-client is usually running without DBUS_SESSION_BUS_ADDRESS set in the process environment. Unfortunately, the "from gi.repository import Gtk" line from epoptes-client python-based utilities autolaunches dbus! This then causes the 1.5 minute shutdown delay issue, because systemd can't cleanly terminate the autospawned dbus process.

To reproduce the issue, boot an LTSP client, don't login, and right click->open a root terminal from epoptes. Verify that ps aux | grep dbus doesn't show a session dbus running. Then send a message to the client, or lock its screen, so that a python utility that imports Gtk runs. Then verify that dbus was autolaunched, wasting memory and blocking clean shutdown.

I've yet to find any clean solutions to this issue, so for the moment I'll be exporting a fake DBUS_SESSION_BUS_ADDRESS so that it's not autolaunched while importing Gtk.

[alkisg]

An additional thing to note here is that dbus isn't killed properly, otherwise it would waste some RAM but it wouldn't delay shutdown.
That might be happening because of this:
To prevent users from killing epoptes, we specify KillSignal=SIGQUIT in epoptes-client.service. Systemd might be sending SIGQUIT to child processes as well, and dbus might be ignoring this.
It will be cleaner if we remove KillSignal=SIGQUIT, and allow the root epoptes-client process to be killed normally with the TERM signal (while still requiring SIGQUIT for the session epoptes-client).

In the future if we use systemd for the session epoptes-client too, we might tell it to keep relaunching and completely remove the SIGQUIT requirement (and even the external socat loop).

[alkisg]

Fix committed in 8361913.

[alkisg]

I'll reopen this as it needs a more brutal workaround:
For example, if an LTSP client is in the login screen, and we Execute sudo gparted in it, then this will be considered a child process of ours, it will spawn dbus, and delay 1.5 minutes on shutdown.

So setting the DBUS_SESSION_BUS_ADDRESS should probably be done in client-functions, not just in our .py scripts.

[alkisg]

So setting the DBUS_SESSION_BUS_ADDRESS should probably be done in client-functions, not just in our .py scripts.

I implemented that in d42e5dd, but I'll keep this issue open until the relaunching logic is reworked and KillSignal=SIGQUIT is removed.