Re: [RFC PATCH] gpg: add support for gpgsm

From: Carlos Martín Nieto <cmn@dwim.me>
Date: Mar 31, 2016, 14:44 UTC
Message-ID: <1459435490.2124.15.camel@dwim.me>
In-Reply-To: <20160331142200.GB31116@sigill.intra.peff.net>

On Thu, 2016-03-31 at 10:22 -0400, Jeff King wrote:

Show 9 quoted lines

> On Thu, Mar 31, 2016 at 03:51:44PM +0200, Carlos Martín Nieto wrote:
> 
> > 
> > Detect the gpgsm block header and run this command instead of gpg.
> This part makes sense to me, and is a strict improvement (though
> offhand, I wonder if any other systems use the generic "BEGIN SIGNED
> MESSAGE" header. The obvious option would be PEM from "openssl
> smime",
> but it is "BEGIN PKCS7").

Yep. It's useful when you can tell what generated the signatures.

Show 9 quoted lines

> 
> > 
> > On the signing side, ask gpgsm if it knows the signing key we're
> > trying
> > to use and fall back to gpg if it does not.
> This part looks like we incur an extra fork/exec each time we sign
> with
> gpg, even if the user doesn't ever want to use gpgsm, or even have it
> installed.

Yes, this is unfortunate as I don't know any other way to tell whether gpgsm (or whatever else) knows about a key. We could try to find gpgsm in PATH but I suspect this would be expensive as well.

> 
> I wonder if there are any hints we can use from the key ident, but I
> suppose not. In the default config, it comes straight from
> $GIT_COMMITTER_*, and is just a name/email.

Both gpg and gpgsm accept any string and try to match it against the information in the keys. So even though gpgsm keys are shown by the program itself with the "0x" prefix, 'gpgsm -k DigiCert' does show me the public key I have for them.

Show 16 quoted lines

> 
> But maybe we could pull this out to a separate config option, like
> "commit.defaultSignatureType", which could be either "gpg", "gpgsm",
> or
> "auto" to enable the behavior you have here.  Then savvy users can
> pick
> the type they plan to use.  We can have a discussion then about
> whether
> to flip the default from "gpg" to "auto", but I'd vote to leave it at
> gpg unless gpgsm gets a huge amount of traction, and it really is
> 50/50
> what people would want.
> 
> And regardless of the default type for creating signatures, we'd
> still
> automatically verify signatures from either type.

This makes sense. It allows for the automatic detection if that's wanted but it would stop us running gpgsm on each invocation.

Show 32 quoted lines

> >  /*
> > + * Try to figure out if the given program contains given the key.
> > Both
> > + * gpg and gpgsm have keys in hex format, so we don't necessarily
> > know
> > + * which one to use.
> > + */
> > +static int program_knows_key(const char *program, const char
> > *signing_key)
> > +{
> > +	struct child_process gpg = CHILD_PROCESS_INIT;
> > +	struct strbuf output = STRBUF_INIT;
> > +	const char *args[4];
> > +	size_t len;
> > +
> > +	gpg.argv = args;
> > +	gpg.in = -1;
> > +	gpg.out = -1;
> > +	args[0] = program;
> > +	args[1] = "-K";
> > +	args[2] = signing_key;
> > +	args[3] = NULL;
> I think you'd want to send stderr to /dev/null here, as this is for
> speculatively seeing "does the user even have gpgsm set up?".
> 
> > 
> > +
> > +	if (start_command(&gpg))
> > +		return error(_("could not run '%s'"), program);
> Likewise, most users would start seeing "could not run 'gpgsm'" if
> they
> do not even have it installed.

Ah yes, I completely forgot to take that into account.

cmn