[pve-devel] [PATCH] run-env: fallback to all zero mac for interfaces without

[Christian Ebner]

On 7/11/25 12:25, Thomas Lamprecht wrote:
> Am 11.07.25 um 12:03 schrieb Christian Ebner:
>> On 7/11/25 11:47, Christoph Heiss wrote:
>>> On Fri Jul 11, 2025 at 11:14 AM CEST, Thomas Lamprecht wrote:
>>>> Am 11.07.25 um 10:27 schrieb Gabriel Goller:
>>> [..]
>>>>> To be honest I'd rather filter out this interface. A zeroed out mac is
>>>>> reserved for loopback interfaces and usually isn't routed.
>>>>
>>>> It's not like we set the MAC to zero, rather it's just used for displaying.
>>>> This way an admin can at least see the interface and select it for usage,
>>>> even if they then need to correctly configure it manually after installation
>>>> to make it actually work.
>>>>
>>>> That said, as manual intervention is required either way, filtering out
>>>> might be OK, but your arguments here are IMO not justifying why that route
>>>> should be chosen. FWIW, a third alternative might be that the rust
>>>> implementation might also just have to learn to not expect a MAC...
>>>
>>> FWIW, there's also been a Bugzilla report a few days about this problem
>>> [0].
>>>
>>> I've took a cursory glance at going about the third route here, although
>>> didn't really get to write much code due to other, more pressing things.
>>>
>>> If anyone wants to pick that up, short summary w.r.t the Rust part:
>>>
>>> - It's mostly about doing a `String` -> `Option<String>` conversion for
>>>     `proxmox_installer_common::setup::Interface`, the MAC address from
>>>     that is then only ever used the post-hook.
>>> - There's also `proxmox_auto_installer::sysinfo::NetdevWithMac`, which
>>>     tries to read the MAC address from /sys/class/net.
>>>
>>> [0] https://bugzilla.proxmox.com/show_bug.cgi?id=6508
>>
>> Okay, can have a look at this too, but do feel free to beat me to it!
>> Not so familiar with the installer codebase and this smells like having
>> some regression potential.
> 
> Yeah, maybe, that's why simply falling back to some "unknown"-like value
> sounded promising to me. If we do not have any parsing/checks for the MAC
> value, we could indeed fall back to a literal "unknown", then it would be
> noticed if we try to use it (e.g., for interface link name-pinning in the
> installer)
> 
> btw. what does the interface looks like, does it really have no MAC
> in the ip link output?

No, the only content for the link line is `link/none`, no mac and brd 
address.