- Date: Tue, 3 Oct 2017 10:32:45 +0100
- From: Simon McVittie <smcv@xxxxxxxxxx>
On Tue, 03 Oct 2017 at 12:12:54 +0530, Pirate Praveen wrote:
> I cannot accept arbitrary interpretations of policy. When build tools
> are not available in main, they cannot go to main, and if the software
> itself is Free Software, it can go to contrib. If you disagree, please
> get the policy clarified. As per the current policy, these packages
> clearly qualified for contrib and these were already accepted by ftp
> masters into the archive.
>From a Freeness perspective, sure, that's contrib.
>From the point of view of security and predictability (which requires
knowing what we're running), it seems undesirable to have a build
process that pulls in arbitrary code from the internet. contrib has
permission to ignore the "all dependencies must be DFSG" clause of Policy
- that's the point of it. However, it does not have permission to ignore
technical/quality-based Policy considerations.
> I thought building these things (making sure the source corresponds to
> the distributed binaries), though using tools outside archive, was
> preferred over shipping pre-built binaries.
If we can't be sure that those out-of-archive tools haven't been
replaced with malware between one build and the next, then I'm not
sure how we can prefer that approach. How can we know that they aren't
writing malicious content into the produced .deb?
Presumably you verified that at the time *you* built the package, the
out-of-archive tools were of a non-malicious version, and were producing
that correspond to the source. It would probably be useful to write down
how you do this in debian/README.source if you haven't already.
However, that verification isn't really sufficient if a rebuild
on the buildds could download an entirely different version of the
out-of-archive tools: a sufficiently inventive attacker who had gained
control over upstream's distribution channel could even arrange to serve
a non-malicious toolchain to your IP address, but then serve a malicious
version to Debian buildds' IP addresses.
(I think downloading versions of the out-of-archive tools that exactly
match cryptographically strong hashes of the same version you used would
be OK from this point of view, although that would be subject to random
build failures if the buildd or upstream has connectivity issues, and
wouldn't work on buildds that block internet access altogether; and if
be feasible to check hashes of everything you execute before executing it.)
> But it seems shipping
> pre-built binaries are preferred. It looks to me like a misplaced
> compromise, but I will follow this advice and ship pre-built binaries
> next time without building them using npm.
At least that way, you have the opportunity to inspect the pre-built
binary (I hope "binary" here means a bundled/minified version that is
not the preferred form for modification but is somewhat human-readable,
rather than something as opaque as a compiled C binary) and have some
level of confidence that it corresponds to the source.
This is the same line of reasoning that says packages like
game-data-packager and flashplugin-nonfree should record known-good
cryptographic hashes of the opaque binary blobs they distribute, even
though there's presumably a lot less ability to audit those than your