Hi,
Adding back SELinux guys (wonder why they got stripped in the initial
posting: I blame the corporate mail client) :(
One question that we have is what is the best way to define a plugin
interface for rpm?
Should we define just security-related hooks and embed currently
existing collection plugin hooks into it
or are there any plans for bigger rpm plugin interface that security
hooks should be part of?
Good question... The current collection hooks aren't really security
at all but something else. Yet the selinux policy plugin does kinda fall
into that category, and if there's an entirely different plugin type needed
for actually putting the security labels (retrieved by eg collection plugin)
and such into place, requiring different plugins to interact somehow, and
that I suppose would get ugly real fast.
There are some other places that could use plugins as well that would fall
into entirely different categories. Just as an example, one such thing is
translations of i18n header strings from external sources (eg specspo which
exists in at least two different flavors, one upstream and one in Fedora)
that could be cleanly handled (and would open up new
possibilities) with plugins.
Yes, I suppose many use cases can benefit of having a genetic rpm plugin
interface, but I guess currently the biggest user of such interface is still
security (with selinux already being in upstream rpm and us dragging patches
around in order to implement our needs).
Just FWIW / food for thought, back some time ago I had a brief look at
isolating the security activity into a relatively abstract security
manager object (instead of plugin hooks littered all around), which could
actually be a stack of different bits and pieces implemented by plugins
(SELinux/MSS labels, maybe file capabilities etc) but that ran into all
sorts of issues right on the first steps already. Whether I just didn't
pursue it hard enough or just not a feasible approach I dunno but suspect
the latter to be the case.
Hm... I think we can make sure that SELinux and our implementation can be
pushed to the security manager. Conceptually I don't see the reason why
pushing capabilities or other default security attributes (like uid, gids
and etc.) to security manager wont' be possible. Could you tell what kind of
issues you have run on the first steps? It isn't going to be an easy and
simple exercise, but should be possible to do.
As for the actual hooks, I think we just need to look at the existing and
planned needs of the security labelling etc stuff one by one and try to
figure whether it all needs to go into a giant transaction plugin
interface or whether it's actually possible to categorize them into tidier
sets.
I think we have these main group of hooks:
- current per-collection hooks
- per transaction hooks (pre-tsm, post-tsm)
- per package hooks (called in psm): pre-psm, post-psm, verify
- per file hooks (called in fsm) + hook for newly created dirs
- script hook (called for any invocation of any script (per-transaction/per
package))
Would you prefer to break them per group or? I am not sure what you meant
with whether it's actually possible to categorize them into tidier sets.?
Some other plugin possibility examples: the different rpmio (de)compression
types would fit the plugin use-case spot-on. Whether having them as plugins
actually makes sense is another question, but it presents some interesting
issues: the rpm file io subsystem lives in librpmio, suggesting the core
plugin infrastructure should be there as well, instead of librpm itself.
However they would need to be able to register themselves into rpmlib()
provides which live in librpm, in the transaction layer...
The ability for plugins to register custom rpmlib() provides seems
generally useful, eg the SELinux/MSSF stuff might want to do it too and I
can imagine other cases as well. And the needs of plugins wanting to
advertise themselves through rpmlib() provides differ vastly - from very
low-level file io to the rather high-level collections and whatnot.
The only thing I know for sure is that there are needs (or at least
opportunities) for more than one plugin type, and the different types
should be able to share the common basic plugin manipulation (loading
etc) infrastructure.
Yes, that's why I was asking what would be the right way to go currently
about the security hooks. I guess it is also can be made step by step: first
unify the security hooks and make sure that all security functionality from
LSMs are isolated to the plugin. And then integrate security plugin
interface to the generic rpm interface. But of course if we can make it
right (to the extend it is even possible) from beginning, it is always
better.
Best Regards,
Elena.
smime.p7s
Description: S/MIME cryptographic signature
___
Rpm-maint mailing list
Rpm-maint@lists.rpm.org
http://lists.rpm.org/mailman/listinfo/rpm-maint
