[libvirt] default virtual network race with NetworkManager in a VM

[Cole Robinson]

Hi all,

There's a long standing issue with running libvirt inside a VM:

https://bugzilla.redhat.com/show_bug.cgi?id=811967

Summary: Run a VM using the default virtual network with IP address
192.168.122.XXX. Install libvirtd inside the VM, reboot. On startup, libvirtd
in the VM can bring up its copy of the default virtual network, before
NetworkManager has a chance to run. The nested network grabs the route for
192.168.122.0 before NetworkManager can claim it for the VM's normal
networking. Result is the VM doesn't have any network connectivity to the
outside world.

If NetworkManager starts first, a libvirt check at virtual network startup
time will refuse to override the existing route and avoid the problem. This is
from https://bugzilla.redhat.com/show_bug.cgi?id=235961

The past few Fedora releases haven't generated that many complaints about
this, but we've had reports that it's happening quite frequently with
rawhide/f21, see the last comments of that first bug report.

There's some suggestions in the bug about possible fixes, like listening to
NetworkManager API events and stopping the virtual network if needed. But they
are just ideas that no one has really researched.

Anyone have a novel solution and/or some cycles to poke at this?

(FWIW no one has tried to determine what has changed that's tickling this
issue more often, but regardless the root issue should be dealt with)

Thanks,
Cole