[pve-devel] [PATCH pve-manager 1/1] cli: add pveeth

[Stefan Hanreich]

On 7/10/25 18:25, Gabriel Goller wrote:
>>> However, unpinning has a problem. Currently, we reset to the
>>> `ID_NET_NAME_PATH` name from `udevadm` and remove the link file, but we
>>> don't know the current systemd `NamePolicy`.

Not sure if users overriding the default naming policy of network
interfaces is a common enough use case to consider here. In that case
users will have to take care of updating / changing the network
interface names themselves anyway, since they would change with that
setting as well.


>>> For example, if the original interface was `ens18` using Slot policy,
>>> after pinning it becomes `nic0`. When unpinning, the interface becomes
>>> `ens18` again (because I'm using slot policy), but the config files
>>> reference `enp0s18` (because `pveeth` assumes we use path policy). This
>>> breaks the network configuration.
>>>
>>> I think the solution is to avoid deleting the link file during
>>> unpinning. Instead, we should just update the interface name.
>>>
>>> The problem is that we can't determine which policy was originally used,
>>> so pin-unpin cycles *don't* restore the original interface name.

But then we'd have changed the name that we're pinning, not removed the
pinning altogether.

Unpin is trying to revert to the name that would be automatically
generated without any link files present, not the exact name that the
NIC had when we generated the pinning. I can see how this might be
obtuse though, particularly since I write that pin / unpin should be
idempotent (which they aren't, because of the behavior I described
above) - my mistake.

With storing names there's still the possibility of the stored name
clashing with another interface in the meanwhile (due to e.g. upgrades).
We cannot just store the old names and reset to them. Even with the same
naming policy, names can change across updates (which is what happens
with the upgrade to systemd v257, for instance), so we have to query how
the interface name would look like with the *current* settings, not with
the settings at the time of pinning. This also applies to storing the
policy, who says that the policy hasn't changed since we stored it?


>> This I'd need to think through, just wanted to comment on above before
>> I forget.
> 
> If we really want to make pin and unpin involutive, we would need to
> store somewhere the interface names or store the interface naming
> policy.

unpin was more intended as a solution for users that made an error with
invoking the pin command and give them an easy way to revert the changes
generated by pin. In the other thread with Dominik I've also discussed a
different approach on how to handle applying the configuration. Solving
it as follows would also introduce a way of reverting the configuration:

* Pinning generates the new configuration files in the pending config of
/e/n/i and SDN. For the firewall we'd have to create one as well and
probably just handle this manually in the following step.

* Add another command that applies the temporary changes which would
also include applying the changes via udevadm immediately.

If we solve it like this, then we could introduce a 'revert' or
'rollback' command, which would simply delete any pending changes and
then remove the generated link files. We'd have three possible actions
for handling pending configuration files:

* generate (generates the pending configuration)
* apply (which applies pending configuration)
* revert/rollback (which removes any pending configuration changes)

This would reset everything to the way it was before generating the
pending configuration. It would also obsolete a dry-run flag imo, since
we have the intermediate, pending, configuration that needs to be
manually applied. Users can use those for inspecting the potential changes.

It would still make sense to provide the opportunity for users to get
rid of all pinned names, which unpin in its current state could then do.