Bug#810367: kmod-static-nodes.service fails without a dependency on kmod

[Felipe Sateler]

Hi Helmut,

On 8 January 2016 at 12:47, Helmut Grohne <helmut at subdivi.de> wrote:
> Source: systemd
> Version: 215-17+deb8u2
>
> Hi Martin, Michael and other systemd folks,
>
> in another attempt at exploring systemd, I noticed that you can install
> a base system with systemd without installing kmod. This is useful for
> minimizing an installation when your kernel does not support module
> loading. It is also useful when systemd's own module loading
> capabilities via libkmod are sufficient.
>
> When doing so however, kmod-static-nodes.service loudly fails finding
> kmod. I argue that this is a bug in the systemd packaging. Either that
> service is optional. In this case it should gain a ConditionPathExists.
> Or it is required for running the system. In which case systemd should
> gain a runtime dependency on kmod. In the former scenario recommending
> kmod still looks like a good idea to me.

kmod-static-nodes.service already has:

ConditionPathExists=/lib/modules/%v/modules.devname

That is, you have a kernel installed that require static device node
creation. On my system I do:

% aptitude why kmod
i   linux-image-amd64          Depends linux-image-3.16.0-4-amd64
i A linux-image-3.16.0-4-amd64 Depends kmod | module-init-tools

(m-i-t is a transitional package for kmod).

So, evidently you have a self-built kernel. Is it reasonable to expect
self-compiled kernel users to satisfy the dependencies of the kernel
packages? I'm not sure there is a clear-cut answer for this. But if
you are compiling a kernel with module support, not installing kmod
looks very weird to me.

>
> In my embedded use case, the service actually creates an empty
> /run/tmpfiles.d/kmod.conf and is thus not necessary for proper
> operation. Thus I have a slight preference for the former solution. At
> the same time I understand that the latter is easier to support.

Or have your kernel not ship the modules.devname if it is empty. (BTW,
is it actually empty, or does it contain a comment? If it is empty,
then maybe we can change to ConditionFileNotEmpty).

>
> I am reporting this bug against the jessie version, because I discovered
> it there, but it applies to the unstable version as well. I do not think
> this is worth fixing in jessie as it is a corner case and has trivial
> workarounds.
>
> Helmut
>
> PS: For better serving the embedded, it would be nice if systemd could
>     provide a smaller installation size (i.e. making more components
>     optional). I understand splitting that systemd in more packages and
>     making some of them optional is considerable amount of work and I
>     have little suggestions on how to do that, but maybe you have. Maybe
>     logind or the some of the dbus support (as dbus is already optional)
>     could be moved somewhere else? Of course the default installation
>     should keep these components.

The largest component is networkd, but it requires a lot of conffile
shuffling, so I'm not sure it is worth the effort. For my own use
cases, I actually want networkd everywhere, so that is not an itch I'm
particularly eager to scratch.




-- 

Saludos,
Felipe Sateler