Re: Upstream dist tarball transparency (was Re: Validating tarballs against git repositories)
Thanks for the response!
On Fri, 5 Apr 2024 11:12:33 +0200, Guillem wrote:
> On Wed, 2024-04-03 at 23:53:56 +0100, James Addison wrote:
> > On Wed, 3 Apr 2024 19:36:33 +0200, Guillem wrote:
> > > On Fri, 2024-03-29 at 23:29:01 -0700, Russ Allbery wrote:
> > > > On 2024-03-29 22:41, Guillem Jover wrote:
> > > > I think with my upstream hat on I'd rather ship a clear manifest (checked
> > > > into Git) that tells distributions which files in the distribution tarball
> > > > are build artifacts, and guarantee that if you delete all of those files,
> > > > the remaining tree should be byte-for-byte identical with the
> > > > corresponding signed Git tag. (In other words, Guillem's suggestion.)
> > > > Then I can continue to ship only one release artifact.
> > >
> > > I've been pondering about this and I think I might have come up with a
> > > protocol that to me (!) seems safe, even against a malicious upstream. And
> > > does not require two tarballs which as you say seems cumbersome, and makes
> > > it harder to explain to users. But I'd like to run this through the list
> > > in case I've missed something obvious.
Ok, after a bit more time to process the details, this makes more sense to me
now. It's a fairly strong assertion about the precise VCS origin and commit
that a _subset_ of the files in a dist tarball originate from.
And the strength of the claim (I think) varies based on how feasible it would
be for an attacker to take control of the origin and write a substitute commit
with the same VCS commit ID and file list - so it's based on fairly
well-understood principles about crytographic hash strength.
(this seems similar in some ways to the existing .dsc file format, although
in relation to a 'source package source' and not sources of binary packages)
In any case: I'm reasonably convinced that the provenance (claim-of-origin)
that this would provide for a source tarball is fairly strong. That's not my
only concern though (in particular, goto: regeneration).
> > Does this cater for situations where part of the preparation of a source
> > tarball involves populating a directory with a list of filenames that
> correspond to hostnames known to the source preparer?
> >
> > If that set of hostnames changes, then regardless of the same source
> > VCS checkout being used, the resulting distribution source tarball could
> > differ.
>
> > Yes, it's a hypothetical example; but given time and attacker patience,
> > someone is motivated to attempt any workaround. In practice the
> > difference could be a directory of hostnames or it could be a bitflag
> > that is part of a macro that is only evaluated under various nested
> > conditions.
>
> I'm not sure whether I've perhaps misunderstood your scenario, but if
> the distributed tarball contains things not present in the VCS, then
> with this proposal those can then be easily removed, which means it
> does not matter much if they differ between same tarball generation
> (I mean it matters in the sense that it's an alarm sign, but it does
> not matter in the sense that you can get at the same state as with a
> clean VCS checkout).
Yep, you managed to translate my baffling scenario description into a clearer
problem statement :)
> The other part then is whether the remaining contents differ from what
> is in the VCS.
>
> If any of these trigger a difference, then that would require manual
> review. That of course does not exempt one from reviewing the VCS, it
> just potentially removes one avenue for smuggling artifacts.
Why not reject an upload automatically if a difference is detected between the
source package source and the dist tarball?
> > To take a leaf from the Reproducible Builds[1] project: to achieve a
> > one-to-one mapping between a set of inputs and an output, you need to
> > record all of the inputs; not only the source code, but also the build
> > environment.
> >
> > I'm not yet convinced that source-as-was-written to distributed-source-tarball
> > is a problem that is any different to that of distributed-source-tarball to
> > built-package. Changes to tooling do, in reality, affect the output of
> > build processes -- and that's usually good, because it allows for
> > performance optimizations. But it also necessitates the inclusion of the
> > toolchain and environment to produce repeatable results.
>
> In this case, the property you'd gain is that you do not need to trust
> the system of the person preparing the distribution tarball, and can
> then regenerate those outputs from (supposedly) good inputs from both
> the distribution tarball, and _your_ (or the distribution) system
> toolchain.
regeneration:
Here is the problem. Let's say that in future with this transparency code in
place, a security bug is discovered in versions of autotools that were
available in the testing or unstable distributions (and may have been used by
some Debian maintainers/developers). It could be useful to determine a number
of things:
* What dist tarballs were built using the affected versions?
* Do those dist tarballs differ when rebuilt with the fixed autotools?
It's similar to the discovery of a security-related problem in a compiler or
other toolchain component; we want to identify and rebuild affected artifacts.
I think your proposal is good at providing provenance for the subset of files
that originate from VCS; what I'm not sure it provides is confirmation about
how to confirm regeneration of the remainder of the dist files.
Another way to think about it is: let's say you _do_ attempt to regenerate a
dist tarball because you're curious -- and the results differ somehow. How
would you find the cause of the difference? Without the source-build-deps,
I'd argue that's nearly intractable in some cases (although probably possible,
to an unreliable degree of trust, by simply re-attempting the build with other
different versions of the build tools). With the source-build-deps provided,
then it should either become trivial (e.g.: oh, I used the wrong version, I'll
rebuild again), or it becomes a question of bug-finding (ok, I used the same
version, but the output differed: that means something here is
non-reproducible and can be considered a bug).
> The distinction I see from the reproducible build effort, is that in
> this case we can just discard some of the inputs and outputs and go
> from original sources.
Ok, understood. I counter-claim that that is _almost_ enough, but that it's
helpful when unexpected problems occur in future to be able to go back and
recreate the preconditions accurately (and when that accuracy isn't enough, to
determine why _that_ is, and resolve that separately).
Regards,
James
Reply to: