bug#40998: Guix System's initrd doesn't honor rootflags

[Ludovic Courtès]

Maxim Cournoyer <maxim.cournoyer@gmail.com> skribis:

> Ludovic Courtès <ludo@gnu.org> writes:

[...]

>> There’s no need to have a ‘version’ field in live <boot-parameters>
>> records: have the ‘version’ field in the serialized format (the sexp)
>> and make sure the deserializer correctly converts to the internal
>> representation.
>>
>> Here, I think you can bump the version number in the serialized form,
>> and have ‘read-boot-parameters’ automatically augment ‘kernel-arguments’
>> when VERSION is 0 with “--root=XYZ”.
>
> I initially went that route, as this was the idea you'd given me
> initially.  However, 'read-boot-parameters' deals with serializing the
> parameters file only; to add 'root', 'gnu.load' and 'gnu.system', the
> operating-system object as well as the root device are needed.

<boot-parameters> already has ‘root-device’, so that’s fine.

But you’re right that the system itself is a problem because that’s
self-referential—it’s the thing the “parameters” file is in.  Hmm!

We could add a substitution mechanism where a literal “$SYSTEM” (say) in
the ‘kernel-arguments’ of <boot-parameters> would be substituted by the
actual system store file name by ‘read-boot-parameters’, but maybe
that’s overkill.

So long story short: keeping the ‘version’ field in <boot-parameters>
sounds reasonable after all.  :-)

> The reason 'gnu.load' and 'gnu.system' aren't written to the
> parameters file to start with is because they would cause the system
> directory to no longer be content-addressable; this is explained in
> the docstring of 'operating-system-boot-parameters-file':
>
>     When SYSTEM-KERNEL-ARGUMENTS? is true, add kernel arguments such as 'root'
>     and 'gnu.load' to the returned file (since the returned file is then 
> usually
>     stored into the content-addressed "system" directory, it's usually not a
>     good idea to give it because the content hash would change by the content 
> hash
>     being stored into the "parameters" file).

This comment originates in 40fad1c24ce60076e26f6dc8096e4716d31d90c3.  I
find it a bit misleading because nothing’s “content-addressed”, but I
guess it refers to the same problem: that this is self-referential.

(There’s only one use of #:system-kernel-arguments? #t.  We can remove
that keyword parameter from ‘operating-system-boot-parameters-file’
since it’s not used there.)

>> Also, you could write the ‘match’ pattern like this:
>>
>>   ('boot-parameters ('version (and version (or 0 1)))
>>                     ('label label) …)
>
> I think this patch's current form is preferable, as it means future
> boot-parameters version bumps won't break older Guices (when
> reconfiguring), as long as the version is an exact, non-negative integer
> (e.g. when going from 1 to 2).

That’s what we want to avoid: bumping the version number means that the
new format is not backwards-compatible, and that older Guix versions
won’t be able to read it.  That’s why I think ‘read-boot-parameters’
needs to be explicit about the version(s) it expects.  (A more complete
example of this pattern is ‘sexp->manifest’.)

Breaking backwards compatibility should be avoided when possible, but
it’s not always possible.  In ‘read-boot-parameters’, ‘bootloader-name’,
‘bootloader-menu-entries’, ‘kernel’, etc. are handled somewhat weirdly
to preserve backwards-compatibility; doing this allowed us to not bump
the file format version.

Thanks!

Ludo’.