bug#22883: Authenticating a Git checkout

[Mike Gerwitz <mtg@gnu.org>]

[-- Attachment #1: Type: text/plain, Size: 4908 bytes --]

Ludo:

On Fri, Jun 03, 2016 at 18:12:47 +0200, Ludovic Courtès wrote:
> First, ‘git pull’ doesn’t do it for you, you have to pass ‘--verify’ and
> there’s no way to set it globally.

That's unfortunate.  Does your checkout scenario include a fresh clone?
If so, a pull flag wouldn't help there.

Leo mentioned a patch; I don't think that'd be too difficult (looking at
other config options in builtin/pull.c), and would be a great idea.  It
appears to pass it off to merge.c; that might be a useful area to verify
signatures as well (pull being a fetch && merge/rebase), in a general
sense.

> Second, even if it did, it would be a shallow check: as Mike notes in
> <https://mikegerwitz.com/papers/git-horror-story> with the ‘signchk’
> script, you actually have to traverse the whole commit history and
> authenticate them one by one.  But that’s OK, it runs in presumably less
> than a minute on a repo the size of Guix’s, and we could also stop at
> signed tags to avoid redundant checks.

Practically speaking, that's probably fine, though note that a signed
tag is just a signed hash of the commit it points to (with some
metadata), so you're trusting the integrity of SHA-1 and nothing
more.

With that said, the tag points to what will hopefully be a signed
commit, so if you verify the signature of the tag _and_ that commit,
that'd be even better.  Git's use of SHA-1 makes cryptographic
assurances difficult/awkward.

An occasional traversal of the entire DAG by, say, a CI script would
provide some pretty good confidence.  I wouldn't say it's necessary for
every pull.

> Third, as I wrote before¹, relying on the OpenPGP web of trust to
> determine whether a commit is “valid” is inappropriate: what we want to
> know is whether a commit was made by an authorized person, not whether
> it was made by someone who happens to have an OpenPGP key directly or
> indirectly certified.

If you want to keep with the convenience of the web of trust, then you
can have a keyring trusting only the appropriate Guix
hackers.  Otherwise, I agree.

> Fourth, there’s inversion of control: ‘git log’ & co. call out to ‘gpg’,
> so if we want to do something different than just ‘gpg --verify’, we
> have to put some other ‘gpg’ script in $PATH.  Blech.

What types of things are you considering?  Or are you just considering
the possibility?

I agree that it is awkward.  At the same time, making it configurable
(in the git sense) can potentially be very dangerous, because a
malicious script (e.g. configure) could just modify it to a noop
(e.g. `true`) and circumvent signature checks.

> Fifth, even if we did that, we’d be stuck parsing the possibly l10n’d
> output of ‘gpg’.  Pretty fragile.

In the log output?  You could use --pretty and %G*.  Otherwise, yes
parsing GPG's output seems dangerous; surely there's a better way (like
Leo mentioned).

> Well no, it turns out that libgit2³ has no support for signed commits
> (the ‘signature’ abstraction there has nothing to do with OpenPGP
> signatures.)

!? D:

That's more concerning from a community mindset standpoint than anything.

> Seventh, even if it did, what would we do with the raw ASCII-armored
> OpenPGP signature?  GPG and GPGME are waaaay too high-level, so we’d
> need to implement OpenPGP (in Guile, maybe based on the OpenPGP library
> in Bigloo?)?!

What about gpgme/libgcrypt?[*]

> I stumbled upon git-lockup⁴, which uses something other than OpenPGP to
> sign objects in Git.  However, signatures are not stored in commits but
> rather in “git notes”, which, IIUC, are mutable objects detached from
> the rest of the object store, so not great.

It seems a bit over-complicated.  Without reading much into it, it
doesn't strike me as much different than a detached signature, but the
problem is that the signature (as you implied) can just be
deleted.  Git's commit/tag signatures are embedded in the actual
object.  git-lockup also seems to hash "(branch,commitid) pairs", which
signs considerably less data than Git's signature would (unless it
actually signs the full object, not a string referencing it).


I'll have to read over your first reference (your message) and its
references; now I'm curious.


[*]: I was actually considering writing an FFI for libgcrypt (if it
doesn't exist already), but it made me uncomfortable without studying
whether Guile can make assurances that pointer-referenced data in
"secure" memory will never be copied anywhere else.  I was going to
bring it up in the near future on the guile mailing list after I did
some research myself; no need to derail the discussion here.

-- 
Mike Gerwitz
Free Software Hacker+Activist | GNU Maintainer & Volunteer
https://mikegerwitz.com
FSF Member #5804 | GPG Key ID: 0x8EE30EAB

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 818 bytes --]