Fix netplan try revert (LP: #2083029)

[MggMuggins]

Description

Fixes LP#2083029

When using systemd-networkd as a backend, netplan try backs up /run/systemd/network before applying the new configuration. If the user fails to confirm the new configuration before the timeout is up, Netplan restores the old configuration by moving the backup into /run/systemd/network and invokes networkctl reload to apply the known-good configuration.

However, networkd will not/cannot read config files if their ownership is not root:systemd-network.

In effect, netplan try does not revert network changes at all when using networkd.

Additionally, following the Ubuntu patches to NetworkManager in 23.10 (see netplan everywhere), netplan generate is called (via DBus) when NetworkManager refreshes its config. This "reverts the revert" from netplan try, so restoring the old config doesn't work with NetworkManager in Ubuntu >23.10 either.

This PR adds an autopkg test for netplan try so as to catch these regressions before they are released.

Implementation Notes

I took advantage of the stamp that exists when netplan try is running. There doesn't appear to me to be a way to prevent NetworkManager from calling generate on reload.

I considered that this may be prone to races if NetworkManager takes its time reloading the configuration; however, it looks like netplan try calls command_apply with sync=True, so it shouldn't delete the stamp until after NetworkManager has finished loading its config.

Test Plan

Using the default netplan configuration in a LXD instance:

network:
    version: 2
    ethernets:
        enp5s0:
            dhcp4: true

Replace the configuration with a static IP and run netplan try --timeout 3, and allow the timeout to elapse.

network:
    version: 2
    ethernets:
        enp5s0:
            dhcp4: false
            addresses:
              - 10.58.215.80/24

Before the fix in Jammy, the static IP remains assigned to enp5s0 after the revert. In Noble and later, the interface has the IPv4 removed, but a new dhcp lease is not acquired.

networkctl status --all will report the device as unmanaged. In Jammy:

● 2: enp5s0
                     Link File: /usr/lib/systemd/network/99-default.link
                  Network File: /run/systemd/network/10-netplan-enp5s0.network
                          Type: ether
                         State: routable (unmanaged)
                  Online state: online

Checklist

• Runs make check successfully.
• Retains code coverage (make check-coverage).
• New/changed keys in YAML format are documented.
• (Optional) Adds example YAML for new feature.
• (Optional) Closes an open bug in Launchpad. LP#2083029

Test failures look like Launchpad having trouble.

[MggMuggins]

I opened this PR prematurely; the current changes do not fix the bug. I would mark this as a draft but I don't have permission to do so. Still working on this...

[MggMuggins]

This is ready for review.

Test failures:

• HTTP request failures (look transient)
• Debian testing autopkgtes

The integration tests pass when run manually in a VM/container (Noble, Oracular, Trixie).

I was able to reproduce the Debian testing failure using autopkgtest locally. I ran the Ubuntu autopkgtests locally; they fail too but with an error message:

test_netplan_try (__main__.TestNetworkManager.test_netplan_try) ... eth43 eth43 Error: NetworkManager is not running.

I'm willing to continue chasing these but I'd appreciate a review of this PR to make sure that this is a reasonable set of fixes before I sink more time into the CI. Thanks!

[slyon]

@MggMuggins Thank you very much for your work on this. I retriggered some of the test cases that failed for odd reasons. The scenarios autopkgtests are now failing for Ubuntu & Debian. This seems related to the netplan try command. Can you please investigate what's happening there?

[MggMuggins]

Yeah, this is definitely a bug in my tests; fixing these is my priority for this morning. Thanks for your patience.

[MggMuggins]

Tests are green.

Also added a commit to fix debug logging of the merged config.

[slyon]

Thank you very much this contribution! This is an excellent analysis and test coverage. Much appreciated! It solves an important issue with netplan try.

I left some minor comments inline. Let me know what you think.

I think this might also solve https://bugs.launchpad.net/netplan/+bug/2027583

[MggMuggins]

Thanks for the review! Test failure should clear on a re-run.

[slyon]

Thank you very much for addressing my remarks. Test are now all green. I just pushed a little rename, as suggested by @MggMuggins (b87c3a5).

LGTM!