Re: [PATCH 0/5] allow creation of users/groups before rootfs creation

[Henning Schild <henning.schild@siemens.com>]

Am Wed, 25 Jan 2023 21:55:00 +0100
schrieb "Schaffner, Tobias (T CED SES-DE)"
<tobias.schaffner@siemens.com>:

> On 25.01.23 17:29, Schild, Henning (T CED SES-DE) wrote:
> > Am Wed, 25 Jan 2023 14:44:40 +0100
> > schrieb Gylstorff Quirin <quirin.gylstorff@siemens.com>:
> >   
> >> On 1/25/23 14:29, Henning Schild wrote:  
> >>> Am Wed, 25 Jan 2023 10:01:51 +0100
> >>> schrieb "T. Schaffner" <tobias.schaffner@siemens.com>:
> >>>      
> >>>> From: Tobias Schaffner <tobias.schaffner@siemens.com>
> >>>>
> >>>> This patch series will allow to specify a `pre` flag for the
> >>>> USER_ and GROUP_ bitbake variables. If this flag is set to
> >>>> `true` the given user or group will be created in the rootfs
> >>>> configuration step instead of on rootfs postprocessing. This is
> >>>> helpful when a specific id should be used which would otherwise
> >>>> be picked by a user or group created by one of the installed
> >>>> packages.  
> >>>
> >>> While i do understand the reason i am not sure how relevant that
> >>> is. Why would anything only function with a fixed ID? Whoever
> >>> provided that thing should maybe fix it.  
> >>
> >> Specific IDs are necessary for Updating an A/B rootfs system with a
> >> persistent partition. In this case you do not want to change any
> >> IDs as this can lead to wrong file permissions.  
> > 
> > I see, that is much more understandable. It also goes into the
> > reproducible build direction.
> > But if that is the case the solution does not seem to be good enough
> > because it only considers users/groups created by isar. Not the ones
> > created by installing debian packages. Where the order could be
> > potentially random. Say you DEBIAN_DEPENDS or IMAGE_PREINSTALL "ftpd
> > wwwd" which will craete users "ftp www" where the two deamons do not
> > have any dep on each other and apt-get could install them in any
> > order. That order might in reality not change too often but it
> > could i.e. when you move from debian11 to debian12 or when you
> > bring the third (or 10th) user-adding package into your new
> > firmware.  
> 
> Thanks for the review!
> 
> This series is not about reproducible builds. The IDs of the users and
> groups that are created by debian packages are expected to change
> between _image_ versions. So we expect these IDs to be different after
> a swupdate.

But your initial argument was that you have to create a user before any
debian package comes, or did i misunderstand? Maybe we need to
differentiate between upstream packages and isar-created ones?

> For reproducible builds it is important that the ordering of installed
> packages and their dependencies stay the same but that is a different
> story. I expect the ordering algorithm for a specific set of packages
> including dependencies to be deterministic but would have to look into
> that in detail.

I expect the algorithm to be non-deterministic ... potentially as you
progress using debian. I can just claim, you have to prove.

> Still, if somebody had a requirement that a user or group of a package
> (e.g. ftp or www) stays the same, one could use this change to create
> the user beforehand and pin its id using this change.

True. But we might still want to keep the old database to run
assertions so such things do not go unnoticed.

> > So what you maybe really want is giving isar an /etc/passwd
> > /etc/group pair. Every new firmware is build with the given layer
> > code and that file-pair from the first release. Where you inject
> > those files between bootstrap and install ... hoping that bootstrap
> > will always be the same. Maybe one can inject before bootstrap ...
> > or write a postproc function that will find all different ids and
> > all files and fix up. Or at least start with an assertion in
> > postproc that looks at the old database.  
> 
> Care that just adding /etc/passwd and /etc/group might not be enough
> but you would have to handle the side effects of a useradd/groupadd
> call properly (E.g. creating the home if set). And I expect more
> things to come up if you have a detailed look.

Yes. Not even want to think about LDAP or yp where the databases live
in different files.

> Take into account that the specific patch that introduces the pre flag
> is small, easy to maintain and configure.

It should still be motivated so that we later understand why we have
it. And i would not call it easy to configure if one does not
understand why ... people will "pre" randomly because they do not know
what they are doing. We moved that stuff from pre to post once, and i
can not help the feeling that people might want to want to mix pre and
post one day ... at which point it might become really un-easy.

> > Is the problem of uid/gid depend on install order known in the
> > debian community and how do others solve it? Gentoo has moved from
> > such dynamic allocation to static some years ago, probably for
> > similar reasons.  
> 
> I am open to discuss this with the reproducible build guys but again
> IMHO this discussion belongs into another thread.

IMHO not, the repro guys want full repro ... which we can happily leave
out when it comes to me. You want partial repro, which we should
discuss here.

I do not want to hold this back, just understand it and see if there
can be a better solution. Feel free to ignore me and merge if nobody
else asks questions.

Henning

> > Henning
> >   
> >>>
> >>> So i am willing to say that this is super-niche! And it deserves a
> >>> niche-solution in its layer, not a feature in Isar.
> >>>
> >>> You could hook in a task between bootstrap and image_install. Or
> >>> you could rebuild a bootstrap package to have reserved ids. You
> >>> could run "the thing" in namespaces ...
> >>>
> >>> So is that really relevant? Please go into detail.
> >>>
> >>> Whatever happens i think the python rewrite is cool. But the code
> >>> may have been coming/inspired from OE ... in which case it would
> >>> not be cool, because it would fork away further.  
> >>
> >>
> >> OE uses a complete different implementation than to original:
> >> https://git.openembedded.org/openembedded-core/tree/meta/classes/useradd.bbclass
> >>
> >> see also:
> >>
> >> https://git.openembedded.org/openembedded-core/tree/meta-skeleton/recipes-skeleton/useradd/useradd-example.bb
> >>
> >>
> >>
> >> Quirin  
> >>>
> >>> Henning
> >>>      
> >>>> A rewrite of the image-account-extension in python was done on
> >>>> the way. This allows us to drop a lot of encoding and parsing
> >>>> code that was used to transition to shell and therefore made it
> >>>> easier to read and maintain.
> >>>>
> >>>> Using python functions for more complex tasks allows us the usage
> >>>> of unittests. A very basic infrastructure for unittesting using
> >>>> the build in python unittest and the bb.parse module was added.
> >>>> This was used to test the re-implementation of the
> >>>> image-account-extension as a first showcase.
> >>>>
> >>>> Tobias Schaffner (5):
> >>>>     simplify image-account-extension
> >>>>     allow creation of users/groups before rootfs creation
> >>>>     create a minimal python unittest infrastructure
> >>>>     add unittests for the image-account-extension
> >>>>     set minimal python version in user_manual to 3.5
> >>>>
> >>>>    doc/user_manual.md                            |   4 +-
> >>>>    meta/classes/image-account-extension.bbclass  | 391
> >>>> +++++++----------- testsuite/unittests/README.md
> >>>> | 28 ++ testsuite/unittests/bitbake.py                |  37 ++
> >>>>    testsuite/unittests/rootfs.py                 |  45 ++
> >>>>    .../unittests/test_image_account_extension.py | 175 ++++++++
> >>>>    6 files changed, 434 insertions(+), 246 deletions(-)
> >>>>    create mode 100644 testsuite/unittests/README.md
> >>>>    create mode 100644 testsuite/unittests/bitbake.py
> >>>>    create mode 100644 testsuite/unittests/rootfs.py
> >>>>    create mode 100644
> >>>> testsuite/unittests/test_image_account_extension.py
> >>>>     
> >>>      
> >>
> >>
> >>
> >>  
> >