Re: [PATCH 1/1] verify-tag/verify-commit should exit unsuccessfully when signature is not trusted

["Karel Kočí" <karel.koci@nic.cz>]

[-- Attachment #1: Type: text/plain, Size: 4411 bytes --]

On Fri, Aug 03, 2018 at 12:06:34PM -0400, Jeff King wrote:
> On Fri, Aug 03, 2018 at 11:43:44AM -0400, Santiago Torres wrote:
> 
> > > This is not a deviation. GPG correctly recognizes difference between trusted,
> > > untrusted and unknown levels. git on the other hand does not. Well it did until
> > > the commit a4cc18f29. That one removed GPG exit code propagation.
> > 
> > Oh wow. Sorry my assumption parted from looking at the code back in the
> > day where this happens. I assumed git was quietly propagating the gpg
> > error code and took it from there. 
> > 
> > Now that I think about it though, verify tag can verify more than one
> > tag. I assume that this would make it difficult to propagate individual
> > errors in trusting. I honestly don't know what's the best way to modify
> > this behavior then.
> 
> I think the only sensible thing is to err on the conservative side, and
> return non-zero if we saw _any_ invalid signature.
> 
> I will note, though, that just checking the exit code of `verify-tag`
> isn't really that thorough. It shows that there was _a_ signature, but
> we don't know:
> 
>   - if it was an identity the user would expect to be signing tags
> 
>   - if it even matches the refname we used to find the tag
> 
> So I'd argue that any real verification needs to either have a human in
> the loop, or implement a custom policy based on reading the full output.
> 
> I know we (and you specifically Santiago) talked about this a while ago,
> and we ended up providing ways to get more information out of
> verify-tag, so that a tool could sit on top of that and implement more
> project-specific policy. I don't know offhand of any reusable tools that
> do so, though.

I think that it would be even legit to exit on first tag verification failure. If
someone wants to really verify all tags then it can be done with simple for loop.
git that way does not have to solve problem of error combination.

>   - if it was an identity the user would expect to be signing tags
That can be done just by using trust levels.

>   - if it even matches the refname we used to find the tag
Can you explain this more? You mean that string (such as v1.1) used to lookup tag
object is not verified as part of that object?

OK I thing that it was enough of abstract concepts from me. Let me explain you
what am I trying to achieve. I am implementing feeds (in other words git
repositories with packages) and package sources verification for OpenWRT. We
(project Turris by CZ.NIC) are signing all our commits and all our tags. Now we
are using small script that is verifying our repositories just before we run
build. That is against keyring maintained on our server. I am trying to extend
that to whole OpenWRT tree. That introduces problem of having a lot of keys and a
lot of packages sharing same allowed keys. Fetching all allowed keys for every
package from key servers is just slow because of that I have to share those
between packages. In general there are two options. First one is to have cache of
already fetched keys in armor format. Second one is to have one keyring and by
setting all keys explicitly as never trusted with package given exception.
Unfortunately first option can't be used because of one other request that is from
our team. We don't want to be forced to update list of allowed contributors to our
projects every time we have new colleague. Solution we come up with is to have
central PGP key that signs our whole team and then verification is done by
allowing GPG to fetch additional keys with max-cert-depth 1. That brings me to git
verify-commit/tag that won't exit with zero code when signature is not trusted.

I have a solution for my problem (calling git verify-* twice and grep). That is
not the point of this email nor this contribution. The point is that although
GPG's behavior of exiting with 0 code when trust level is unknown is unexpected
but in the end understandable, git's behavior of exiting with 0 code even if key
is explicitly untrusted is just counterintuitive. I think that few people are
still going to get nasty surprise when I consider that this change was introduced
mid 2014 just after v2.4.0 and Ubuntu 14.04 lts (running even on part of our
infrastructure) still contains version 1.9.1 and in that release it was
acknowledging GPG exit code.

K.K.

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 833 bytes --]