[systemd-devel] Storing package metadata in ELF objects

[Guillem Jover]

On Fri, 2021-05-14 at 14:41:32 +0100, Luca Boccassi wrote:
> On Fri, 2021-05-14 at 12:41 +0200, Guillem Jover wrote:
> > On Sat, 2021-04-10 at 13:38:31 +0100, Luca Boccassi wrote:
> > > On Sat, 2021-04-10 at 13:29 +0100, Luca Boccassi wrote:
> > > > After an initial discussion [0], recently we have been working on a new
> > > > specification [0] to encode rich package-level metadata inside ELF
> > > > objects, so that it can be included automatically in generated coredump
> > > > files. The prototype to parse this in systemd-coredump and store the
> > > > information in systemd-journal is ready for testing and merged
> > > > upstream. We are now seeking further comments/opinions/suggestions, as
> > > > we have a few months before the next release and thus there's plenty of
> > > > time to make incompatible changes to the format and implementation, if
> > > > required.
> > 
> > I've skimmed over the discussion at [0], and while having this data
> > seems like it might be "nice", I've to agree with the comments there
> > voicing that there does not really seem to be an actual need and the
> > overhead and additional work do not seem worth it, TBH, at least
> > in the Debian context.

> Hi Guillem, thanks for having a look, much appreciated!
> 
> Just to clarify, the need is there - this is not an experimental
> exercise, but it is borne out of an actual need&requirement, and it is
> undergoing testing right now before deployment in a large scale
> production infrastructure.
> Not _everybody_ will need it, and not everywhere - that's absolutely
> fair, and discussions on whether the ovearhead is worth it for
> something that is not universally needed, but only in certain use
> cases, is perfectly reasonable and welcome. I know Zbigniew is going to
> try and get some raw numbers on the kind of overhead we are talking
> about, that will hopefully help frame the discussion with more
> precision.

Sorry, I think I expressed myself sloppily. My impression is not
that there's no need behind this, but that there's no need for this
specific implementation at least in the Debian context, as I think
the information should be already available by other means.

Also when I mentioned overhead here, it was more about packaging and
integration work than say actual size increase, even though that might
have an adverse effect on .udeb's for example, and perhaps a
non-insignificant one on .deb's depending on the amount of objects
included.

> > > > The Fedora Wiki and the systemd.io document have more details, but to
> > > > make a long story short, a new .notes.package section with a JSON
> > > > payload will be included in ELF objects, encoding various package-
> > > > build-time information like distro name&version, package name&version,
> > > > etc.
> > > > 
> > > > To summarize from the discussion, the main reasons why we believe this
> > > > is useful are as following:
> > > > 
> > > > 1) minimal containers: the rpm database is not installed in the
> > > > containers. The information about build-ids needs to be stored
> > > > externally, so package name information is not available immediately,
> > > > but only after offline processing. The new note doesn't depend on the
> > > > rpm db in any way.
> > 
> > In the Debian context, the build-ids data is going to be available
> > in the affected executables, and in debug symbols packages and the
> > Packages metaindices listing them, so there's no need for access to
> > any local dpkg database. Given that someone needing to install debug
> > packages will need access to those indices (either with outgoing network
> > access or with a repository mirror), these can be queried at that time.
> > Not to mention that AFAIR the Debian debug symbol servers make it
> > possible to query for specific build-ids.
> 
> This is not strictly related to debug packages, though?

That was actually the impression I was getting from reading the refs
though. :) But I'll expand below.

> In fact, on
> systems where this could be of most use you explicitly do _not_ install
> debug packages (or anything at all). Or even if you wanted to, you
> could not - corefiles are not handled inside the container, but
> outside. Even if you wanted to and were allowed to (which for many
> environments it's not the case), you can't install a Debian debug
> package on a CoreOS host or Mariner host or a Flatcar host.

Sure, but precisely dbgsym .deb's are the things that are trivial to
unpack with dpkg-deb (or ar+tar) w/o needing to actually install them.

> > > > 2) handling of a core from a container, where the container and host
> > > > have different distros
> > 
> > How each distribution handles debug packages and debug symbols is
> > going to be different, so it seems there will be a need for explicit
> > handling of these, at which point the above mentioned querying can be
> > implemented as well, w/o the need to encode the packaging data inside
> > the executable.
> 
> Again, matching to debug symbols is not the main goal here, build-id
> works for that. The main goal is to have useful metadata immediately
> available in all occasions, regardless of where the core was generated
> on the host, without reaching out to external services, so that it is
> directly included and collated in the system journal when the core file
> is handled.
> 
> With a common metadata definition, there's no need to query or
> explicitly handle anything - this already works if you use systemd-
> coredump built from the main branch, and handle a core file from
> different containers running different distros with binaries having
> this metadata in the ELF file, and it just works. This is tested, not
> theoretical.

So this is where I guess I'm missing something. To be able to make
sense of the coredumps there are two things that might end up being
relevant, backtraces and source code. systemd-coredump might already
emit a backtrace, and depending on the information provided it might
be more or less useful. If one needs the actual debug symbols there's
already some external querying/fetching required, and if distribution
specific source code is required because many distributions patch
upstream source, then even more querying/fetching will be required.

Which is why I'm not seeing why this standalone and isolated metadata
would be of much help by itself. As in, the way I see it, either the
information from systemd (w/o the extra metadata) is sufficient to
track down bugs, or that querying/fetching would be needed anyway, at
which point the metadata can be inferred too then?

> > > > 3) self-built and external packages: unless a lot of care is taken to
> > > > keep access to the debuginfo packages, this information may be lost.
> > > > The new note is available even if the repository metadata gets lost.
> > > > Users can easily provide equivalent information in a format that makes
> > > > sense in their own environment. It should work even when rpms and debs
> > > > and other formats are mixed, e.g. during container image creation.
> > 
> > I'm not sure I see the problem here. Either these self-built 3rd-party
> > packages are kept in repos that also provide the debug symbols
> > somewhere for all historically released versions or these will not be
> > accessible anyway. If they are, they can as well be located as per
> > above from the Packages metaindices, and even if the repository
> > metadata gets lost, as long as the debug symbol packages are present
> > (if they are not what's the point anyway) the build-ids can always be
> > re-scanned from them as they are part of the Build-Ids field in the
> > .deb control file.
> 
> I think you are thinking about being confined to single distro here - I
> am talking about mixed environments. Maybe the build-id can be used to
> trace it back - maybe it is available on some federated server. Most
> likely it is not, in these cases - it would require customers to go and
> upload their sources and symbols and information to random third party
> services, a bit unlikely. You have no reference that indicates where to
> go look for. Just a filename, which can be anything, and a random id
> which might never be published anywhere. What do you do with it?

Oh, I was thinking about those mixed environments, full chroots or
stripped down containers, from different vendors, but affecting Debian
installations. What I'm also probably missing here is how does the
metadata help for a third-party that is not expected to track/keep/upload
debug symbols nor source packages into some repository, because otherwise
I'm not seeing how they'd make use of the cores (if they are insufficient
by themselves) to debug issues? I guess my question back would be what
would they do with the metadata if they do not have the debug symbols
nor the sources readily available? Also assuming of course they exercise
good practices such as never reusing the same package-version-arch tuple
for different builds, and similar. :)

> > > > Other than in Fedora, we are already making the required code changes
> > > > at Microsoft to use the same format&specification for internally-built
> > > > binaries, and for tools that parse core files and logs.
> > > > 
> > > > Tools for RPM and DEB (debhelper) integration are also available [3].
> > 
> > So, to conclude, I don't really see the point of this in the Debian
> > context. (Not to mention the problems with encoding binary versions
> > that might be wrong, and the busy work involved.)
> 
> Sorry, what do you mean here by encoding binary versions?

(Expanding, although I assume most of this you already know, but just
for context's sake.)

As was mentioned on the systemd issue (AFAIR), the binary version for
a .deb is decided by the packaging and can (and does) vary from the
source version, and that information is generated after the programs
have been compiled.

For a source version 2.0-1, there are two types of diverging binary
versions, one is for binary-only non-maintainer-uploads (binNMU), such
as 2.0-1+b3 (this would be known at compile-time), and the other would
be for a different binary version-space such as 3:4.5-1 (which is
unknown at compile-time). Both can end up being combined too, as in
3:4.5-1+b5.

Thanks,
Guillem