Re: Bug: machine id is never generated

[Gylstorff Quirin <quirin.gylstorff@siemens.com>]

On 7/22/22 11:32, Gylstorff Quirin wrote:
> 
> 
> On 7/21/22 17:41, Moessbauer, Felix (T CED SES-DE) wrote:
>> Hi,
>>
>> when booting plain ISAR images (with Debian11), the "/etc/machine-id" 
>> is never generated.
>> This breaks a couple of services that depend on having the id.
>> An example is the systemd-networkd with DHCP, leading to error 
>> messages like this one (and breaking networking):
>>
>> systemd-networkd[277]: enp8s0: DHCP6 CLIENT: Failed to set identifier: 
>> No such file or directory
>>
>> The error can manually be fixed by running "systemd-machine-id-setup", 
>> but this obviously does not work in embedded scenarios.
>>
>> The root cause could be that /etc is read-only mounted when the 
>> first-boot-complete.target is reached.
>> At least the logs indicate this:
>>
>> journalctl --grep machine --no-pager
>> -- Journal begins at Thu 2022-07-21 15:07:21 UTC, ends at Thu 
>> 2022-07-21 15:35:20 UTC. --
>> Jul 21 15:07:21 test-image systemd[1]: System cannot boot: Missing 
>> /etc/machine-id and /etc is mounted read-only.
>> Jul 21 15:07:21 test-image systemd[1]: 1) /etc/machine-id exists and 
>> is populated.
>> Jul 21 15:07:21 test-image systemd[1]: 2) /etc/machine-id exists and 
>> is empty.
>> Jul 21 15:07:21 test-image systemd[1]: 3) /etc/machine-id is missing 
>> and /etc is writable.
>> Jul 21 15:07:24 test-image systemd[1]: Condition check resulted in 
>> Commit a transient machine-id on disk being skipped.
> 
> /etc is read only after the initrd gives control to systemd. The remount 
> from ro to rw of the rootfs occurs in systemd.
> 
> 
> I tested it with my Debian machine and without /etc/machine-id to same 
> behavior occurs.
> 
>  From my testing setting '/etc/machine-id' to "unitialized\n" results in 
> the intended behavior with active first boot mechanic.
> 
> 
> ```
> modified   meta/classes/image-postproc-extension.bbclass
> @@ -57,7 +57,7 @@ ROOTFS_POSTPROCESS_COMMAND =+ 
> "image_postprocess_machine_id"
>   image_postprocess_machine_id() {
>       # systemd(1) takes care of recreating the machine-id on first boot
>       sudo rm -f '${IMAGE_ROOTFS}/var/lib/dbus/machine-id'
> -    sudo rm -f '${IMAGE_ROOTFS}/etc/machine-id'
> +    echo  "uninitialized\n"  | sudo tee '${IMAGE_ROOTFS}/etc/machine-id'
>   }
> ```
> 
> This still needs to be tested with other images.

We also tested it with Fedora 36 (thanks to Florian) and if you you have 
no `/etc/machine-id` systemd also fails.

Setting the content of /etc/machineid to "unitialized\n" leads to the 
intended  behavior.

Quirin

>>
>> Maybe Quirin knows more, as he implemented the postproc removal in 
>> 8b5e3f9.
>> IIRC Jan also mentioned something that "rw" has to be added to the 
>> kernel cmdline, but this looks like a workaround as well.
> 
> 
>> For the sake of completeness:
>>
>> cat /proc/cmdline
> 
>> initrd=\initrd.img-5.10.0-16-amd64 LABEL=Boot 
>> root=PARTUUID=9ad286e9-8df8-42c2-ac7d-0f2d7387d03d rootwait 
>> console=tty0 console=ttyS0,115200
>>
>> IMO this is a pretty critical bug as it affects a ton of images.
>>
>> Best regards,
>> Felix
> 
>