Thanks Cyphermox for co-testign this with me.
Here an example of his better interface overview now:

virsh # iface-list
 Name                 State      MAC Address
---------------------------------------------------
 enp3s0               active     c8:60:00:6d:8c:07
 lxdbr0               active     fe:34:73:b4:77:ab
 maas                 active     c8:60:00:6d:8c:07
 staging              active     c8:60:00:6d:8c:07

Vlans are not always listed, but that still is much better than before.

FYI: there is a test build in https://launchpad.net/~ci-train-ppa-service/+archive/ubuntu/3251

I did my own tests on this as well, on an upgrade from xenial all that most have in their E/N/I is just "lo" so that is what they see (but nobody wants/needs lo - it is just there because it is listed in E/N/I).

$ virsh iface-list
 Name                 State      MAC Address
---------------------------------------------------
 lo                   active     00:00:00:00:00:00

After the update to the proposed version I got much better output.
$ virsh iface-list
 Name                 State      MAC Address
---------------------------------------------------
 enp0s25              active     54:ee:75:61:c1:97
 lxdbr0               active     fe:24:e7:8e:21:8d
 wlp4s0               active     94:65:9c:0e:35:12

Btw those are just the active interfaces, you can list passive/inactive ones still (e.g. virt manager shows these greyed out).

$ virsh iface-list --all
 Name                 State      MAC Address
---------------------------------------------------
 conjureup0           inactive   12:ab:dc:f1:41:bd
 enp0s25              active     54:ee:75:61:c1:97
 lo                   inactive   00:00:00:00:00:00
 lxdbr0               active     fe:24:e7:8e:21:8d
 strswanbr1           inactive   52:54:00:fc:52:4d
 strswanbr2           inactive   52:54:00:14:9e:81
 virbr0               inactive   52:54:00:f4:ea:12
 wlp4s0               active     94:65:9c:0e:35:12

I even see slaves devices on bridges in virt-manager now.

Further functions of the iface- API space are good as well, like name<->mac translation.
$ virsh iface-name 54:ee:75:61:c1:97
enp0s25
$ virsh iface-mac enp0s25
54:ee:75:61:c1:97

Even XML generation works if you want to use the snippet for other defines (e.g. guest forwards).
$ virsh iface-dumpxml enp0s25
<interface type='ethernet' name='enp0s25'>
  <mtu size='1500'/>
  <link speed='1000' state='up'/>
  <mac address='54:ee:75:61:c1:97'/>
</interface>

Only the define/destroy/edit actions are now (intentionally) blocked.
The console calls are ok in the refusal message, just like in virt-manager.
$ virsh iface-edit enp0s25
error: this function is not supported by the connection driver: virInterfaceDefineXML
Failed. Try again? [y,n,f,?]: 

From a "virt-manager" and similar tools perspective making changes now looks reasonably guarded.
If you try to do so you get "Error setting ... this function is not supported by the connection driver: virtInterdaceDefineXML"

I also wanted to check a few more corner cases, so I did:
- s390x system (for having odd device types) with some defined in E/N/I (was a Xenial upgrade)
- a lxd container for udev/containers can be odd
- Thanks to cycphermox for his MAAS based tests as well

The container looks good, from an empty device view it got to
# virsh iface-list --all
 Name                 State      MAC Address
---------------------------------------------------
 eth0                 active     00:16:3e:35:78:53
 lo                   inactive   00:00:00:00:00:00
 virbr0               inactive   52:54:00:ec:18:e5
Which ist just matching my "ip link" output

The s390x system had shown the E/N/I devices before upgrade to the fixed version for being found in E/N/I but missed a few others.

$ virsh iface-list --all
 Name                 State      MAC Address
---------------------------------------------------
 encc000              active     02:00:00:8e:1a:31
 encc000.2654         active     02:00:00:8e:1a:31
 lo                   active     00:00:00:00:00:00

With udev mode this is much better catching all the others on my system that were formerly forgotten.
$ virsh iface-list --all
 Name                 State      MAC Address
---------------------------------------------------
 encc000              active     02:00:00:8e:1a:31
 encc000.2654         active     02:00:00:8e:1a:31
 encc003              inactive   02:00:00:28:29:a6
 lo                   inactive   00:00:00:00:00:00
 lxdbr0               active     fe:42:16:55:ae:8b
 ovs-system           inactive   4a:fd:83:ed:05:89
 ovsbr0               inactive   52:17:63:0d:4b:ab
 virbr0               inactive   52:54:00:6e:c3:70


So all that seems good to me to go forward with that approach - and even worst case if we identify issues down the road it seems much better than the current broken state.

Furthermore we actually don't want libvirt to try to be another netplan that translates XMLs into different backends. There is a reason why this was broken so long and never anybody complained - this particular subfeature of libvirt is mostly unused.
I discussed with several people and the general reply was "what having a remote to libvirt allows them to change host interfaces, I don't want that ...".
So lets disable the write support for now and use udev to provide a fixed interface listing.


I started a set of generic regression tests on x86/ppc64el/s390x against this before pushing to cosmic. I'll report on test results later, but in general we want to sync on details like:
- ack to push this to cosmic
- ack and coordinate on an SRU for this to Bionic
- Do we want to archive-remove netcf in Cosmic (or give it a ifupdown dependency)?
- netcf in Bionic remove/dep/..?