Re: [pve-devel] [RFC PATCH common] SysFSTools: mdev: retrieve Nvidia vGPU description from nvidia-smi

[Christoph Heiss <c.heiss@proxmox.com>]

Thanks for chiming in!

On Wed, Oct 30, 2024 at 01:08:28PM GMT, Thomas Lamprecht wrote:
> higher level comment, so CCing also Dominik:
>
> Is pve-common, and possibly even SysFSTools the right place for this?

To be honest, that was kinda my thought to when I started implementing
this, but didn't know quite where to put it otherwise - other than a
completely new package and whether that is worth the additional effort.

So no, not really.

> [..]
> from recent development it seems like placing such things into a more
> specific package, and probably also (binary) package might be better.
>
> So while you certainly do not need to fix all of that just to get your
> changes here in, it would be OTOH great if we could not make this tech
> debt worse; so lets create a new package and possibly also finer-grained
> modules for this.

If that's the way we want to go tho - I'd be happy to put something
concrete together to get started. This change is be a good starting
point for all of this anyway, I think.

And if we split it out anyway, I would really go the way of directly
using libnvml to retrieve all the information. It is not that much more
effort overall and since it being a binary package is fine too, seems
fitting.

>
> The hard thing is to carve out what belongs together, and lets not be
> strictly limited by existing module layout, method from those can be
> split and merged.
>
> Some rough/unfinished ideas/proposal from top of my head:
>
> - possible package names:
>   - libpve-device-common (or host-device / hw-device as slight alternations)

`libpve-host-device(-common)` sounds pretty good IMO; in that it is
quite distinctive about what devices the package actually works with.

`device` alone would be a bit ambiguous, in my mind.

In the long-term, I guess the mdev-related stuff could be moved there
too, being quite a good fit for that too.

>   - libpve-sysfs-common (if sysfs is really only what this does, which then
>     should not need any external tools or compiled programs)
>   - libpve-hw-passthrough-common
>
> - modules: The basename might depend a bit on the package named chosen, could
>   be e.g. "PVE::Device" or "PVE::SysFS"
>   - $basename::USB
>   - $basename::PCI
>   - $basename::PCI::NVIDIA

.. and splitting up such things into vendor specific modules further
down is a good idea too. Keeps things together that should be.

>
> In general, it can be also fine to have a very generic (micro) package with
> just sysfs helpers, or keep that in existing pve-common, and then depend
> on that in a more specific package that provides passthrough related stuff
> on top of that, mixing sysfs query/writing with some potential calling of
> external tools.
>
> As said, the direction should be that, and great if some parts of pve-common
> can be improved "for free" w.r.t. not being huge modules in a huge package,
> but there's no need to rework all of pve-common now already just for this.


_______________________________________________
pve-devel mailing list
pve-devel@lists.proxmox.com
https://lists.proxmox.com/cgi-bin/mailman/listinfo/pve-devel