Re: [PATCH 0/3] credential: handle partial URLs in config settings again

From: Johannes Schindelin <johannes.schindelin@gmx.de>
Date: Apr 22, 2020, 22:26 UTC
Message-ID: <nycvar.QRO.7.76.6.2004230019170.18039@tvgsbejvaqbjf.bet>
In-Reply-To: <20200422221334.GA571740@coredump.intra.peff.net>

Hi Peff,

On Wed, 22 Apr 2020, Jeff King wrote:

Show 22 quoted lines

> On Wed, Apr 22, 2020 at 08:51:02PM +0000, Johannes Schindelin via GitGitGadget wrote:
>
> > This fixes the problem illustrated by Peff's example
> > [https://lore.kernel.org/git/20200422040644.GC3559880@coredump.intra.peff.net/]
> > , in maint-2.17:
> >
> >   $ echo url=https://example.com |
> >     git -c credential.example.com.username=foo credential fill
> >   warning: url has no scheme: example.com
> >   fatal: credential url cannot be parsed: example.com
> >
> > The fix is necessarily different than what was proposed by brian
> > [https://lore.kernel.org/git/20200422012344.2051103-1-sandals@crustytoothpaste.net/]
> > because that fix targets v2.26.x which has 46fd7b390034 (credential: allow
> > wildcard patterns when matching config, 2020-02-20).
> >
> > This patch series targets maint-2.17 instead (and might actually not be able
> > to fix maint due to that wildcard pattern patch; I haven't had the time to
> > check yet).
>
> Yes, this will remove the die() in all versions, but in v2.26.0 and up,
> that config would be silently ignored (and your new test will fail).

Thanks for testing!

Show 16 quoted lines

> > Please note that Git v2.17.4 will not do what we would expect here: if any
> > host name (without protocol) is specified, e.g. -c
> > credential.golli.wog.username = boo, it will actually ignore the host name.
> > That is, this will populate the username:
> >
> >   $ echo url=https://example.com |
> >     git -c credential.totally.bog.us.username=foo credential fill
>
> That seems scary. What if it is not .username, but:
>
>   [credential "example.com"]
>   username = foo
>   helper = "!echo password=bar"
>
> ? (Or you can imagine a helper that is pulling from a read-only store,
> like "pass"). That would send the credential to the wrong host.

It would. But I am not aware of any implications regarding `.gitmodules` (for some reason I now expect every bug to open an attack vector via submodules, I wonder why that is), so that's at least good.

There do seem to be a few projects that set a `credential.<hostname>.useHTTPPath` as part of their build, but I doubt that that could be exploited.

> I think any fix we do here needs to make sure we are not any
> reintroducing wrong-host problems (though I do admit that the usage like
> my example above is probably way less common than vanilla helpers that
> do the host-selection themselves).

Yes. For the record, I tried to be very careful here. The _only_ code path that is affected by this change is the config reading. Reading `.gitmodules` (also in `fsck.c`) should behave identically to the way v2.17.5 behaves: they should not allow empty protocol or hostname. If you spot any difference in behavior in this regard, please do let me know: that would be a bug.

And I do not think that these patches could re-introduce the problems you fixed: the `credential_match()` tries to work against a `struct credential` that has been sent to `credential_apply_config()` and as per your fixes, this function now verifies that neither protocol nor hostname are unset.

But again, I would love another pair of eyes on this, to confirm my analysis.

Thanks, Dscho