Bug#1063881: nvidia-graphics-drivers: provide dependency package to catch all packages of given version

[Drew Parsons]

On 2024-02-15 14:20, Andreas Beckmann wrote:
> On 14/02/2024 00.37, Drew Parsons wrote:
>> It would be much easier to switch between versions in unstable and
>> experimental, or upgrade from experimental, if there were a dummy
>> dependency package that depends on all the manifold nvidia component
>> packages for the given version.  It could be called nvidia-driver-all,
>> for instance.  Then only the one package needs to be marked for
>> upgrade (or downgrade) and will bring in all the others.
>> 
>> Can it be done?
> 
> Yes. I'd probably call it nvidia-driver-full (as in texlive-full)
> since -all could be mistaken as 'installs all (supported) driver
> series'.

That sounds sensible.

> And you would want hard Depends and no Recommends ?

I think it would need to be a hard Depends.  Otherwise a Recommends 
would only activate once the first time the dependency package is 
installed. Since it's not mandatory, it wouldn't succeed in maintaining 
consistent versions when upgrading or downgrading.  A Recommends (=) 
together with a Conflicts would not work since the versioned 
dependencies don't have a != operator to use with Conflicts.

> Is there anything that should be excluded?

Only question I can think of for exclusion is whether cuda should be 
included. For sure not everyone who would want the driver upgrade would 
necessarily want cuda as well, in the sense that they simply aren't 
using cuda.   So one option is make two dependency packages, 
nvidia-driver-full for the drivers without cuda, and nvidia-cuda-full 
(or just cuda-full) for the cuda components.  I guess nvidia-opencl-icd 
(nvidia-opencl-common) might belong in nvidia-driver-full since it's 
kind of a "conflict of interest" to put it with cuda.

Two dependency packages like this would meet requirements fine I think.  
But if it's too much trouble to split them that way and you'd prefer one 
dependency package, then I'd suggest including the cuda packages in it.

> Are there any binary packages from different source packages that
> should be included as well? Mainly thinking about bits that are
> included in the .run file but since source is available, we build it
> from source instead. nvidia-settings, nvidia-xconfig,
> nvidia-persistenced?

I don't think the dependency package would need to set external 
dependencies.  The actual binary packages would bring these in as needed 
in their own Dependencies.  The dependency package would just need to 
make sure all the nvidia package versions remain in step.

Drew