in particular: have subiquitycore.netplan understand them. The goal of
all this is to make the recent change to not delete virtual interfaces
if their configuration has not changed actually work.
Normally netplan.Config().config is a dict as produced by
yaml.safe_load(). However, on systems without any netplan configs that
is not true. The subiquity code that calls into this function expects
.config to be a dictionary. When it's not
subiquitycore/models/network.py produces traceback that "list does not
have get method" when trying to get 'network' key.
The main thrust of this is to not create virtual interfaces until
applying the config.
This meant that the network model has to change a bit to be able to
represent interfaces that do not yet exist on the system. I did this
by ripping out most of the existing network device code: now a
NetworkDev is really just a wrapper for the config for a device and (if
it exists) the netlink data too. A few places had to adjust to checking
if the netlink info is available before accessing it but all in all it
was not that painful.
There are a few other refactorings in this commit that perhaps should be
split out (how the bond parameters are handled, some stuff about
resizing the table rows when interfaces are edited) but it doesn't
really seem worth it.
When user configures network with subiquity, it's rendered
netplan should be wholly definitive. So, we remove the other files
that may have config. This fixes a bug where running in an instance
when running on a system where cloud-init had rendered a 'match' with
'macaddress'.
When writing netplan we keep 'macaddress' match in place but drop
others. The others may just wildcard from the installer environment,
but macaddress are likely by cloud-init or otherwise intentionally
written.
Also add an atomic write in subiquitycore/file_util and move the
netplan code into subiquitycore/netplan.py, and add some unit test
helpers from cloud-init.
Michael Hudson-Doyle