Re: [TLS] Server-side missing_extension MUSTs

[Hubert Kario <[e-mail]>]

On Wednesday 13 July 2016 14:43:58 David Benjamin wrote:
> To be clear, I am not at all opposed to useful errors or strict policing of
> what the peer sends. That's all great. Indeed the linked test below makes
> use of more specific error codes than TLS provides. But I would like such
> things to:
> 
> 1. Be useful, either for debugging or because it rejects an invalid
> handshake that would otherwise go unnoticed.
> 2. Come out naturally out of what the implementation would do anyway, to
> keep complexity down.
> 
> I'm also quite okay with budging on #2 where #1 is strong. Complexity is
> the currency we pay for adding things. But I do not think this particular
> error code passes #1 (the special-case is not useful and such a peer would
> not be able to handshake against correct implementations anyway), so I'm
> not willing to sacrifice #2 for it.

Then I fail to see how missing_extension does not flow out of correct
implementation (and from general RFC recommendations).

You need to parse the Client Hello, and then sanity check it (length of
session ID, validity of ciphers, lack of duplicate extensions, etc.).
Making sure that mandatory extensions are in place in the same code
seems to me like a obvious place to do it.

In other words, you shouldn't delay checking if a particular extension
is present until the time you want to use it. If there is any chance
(or possible configuration) in which you would end up using it, you
should complain that the extension is missing or it is malformed.

This way when you start to depend on it, things won't start breaking
suddenly and you end up writing Yet Another Workaround for Broken
Implementations™
 
> On Wed, Jul 13, 2016 at 10:02 AM David Benjamin <[e-mail]>
> wrote:
> 
> > On Wed, Jul 13, 2016 at 3:31 AM Dave Garrett <[e-mail]>
> > wrote:
> >
> >> To be clear though, completely mandatory extensions, at least as we're
> >> doing them here, are a new thing. TLS 1.2 relied on separate messages for
> >> stuff, but we're front-loading everything into the ClientHello to get
> >> reliable 1RTT (with the exception of HRR).
> >
> >
> > On Wed, Jul 13, 2016 at 8:12 AM Hubert Kario <[e-mail]> wrote:
> >
> >> On Wednesday 13 July 2016 05:23:53 David Benjamin wrote:
> >> > I don't believe an implementation which fails to send supported_groups,
> >> > etc., in 1.3 would ever leave a developer's workstation. It would not
> >> > interoperate with anything.
> >>
> >> it would interoperate with itself, and for some deployments that's enough
> >> of a passing grade... (Even if you do interoperatbility testing you
> >> do not check all possible permutations of features and settings)
> >>
> >> I wholeheartedly agree with Dave here, error definitions should be strict
> >> (both on the when and what to do). One, because it allows to better
> >> diagnose (in general, maybe not in this specific situation) problems.
> >> Two, because you can write a strict negative test case that actually
> >> checks for it.
> >
> >
> >
> > https://boringssl.googlesource.com/boringssl/+/1c256544dda26e4042c1af082580a1b87c9a690f/ssl/test/runner/runner.go#5646
> > Also something I have plenty of experience with. We're obsessive about
> > adding this kind of test in BoringSSL[0]. :-)
> >
> > One doesn't need this alert to write a test for curve negotiation. Just
> > test that the handshake hits the usual codepath for there not being a
> > common cipher.
> >
> > This semi-mandatory extension isn't new in TLS 1.3, depending on your 1.2
> > behavior. BoringSSL will already refuse to select an ECDHE cipher if
> > supported_curves is missing. 1.2 does not require this, but we opted to do
> > it for simplicity[1].
> >
> > David
> >
> > [0] If anyone wants to try, I'm sure there is an implementation-agnostic
> > TLS protocol test suite one could extract out of that. It would take some
> > wrangling as we cheat and condition on BoringSSL error codes, but the nice
> > thing is I've already built up a large battery of tests here.
> >
> > [1] Previously we always picking our most preferred curve if the extension
> > was missing, like OpenSSL. But now our most preferred curve is X25519, not
> > P-256. It did not seem worth kludging that when we could just decline the
> > cipher. No one's noticed, so I would say that worked.
> >

-- 
Regards,
Hubert Kario
Senior Quality Engineer, QE BaseOS Security team
Web: www.cz.redhat.com
Red Hat Czech s.r.o., Purkyňova 99/71, 612 45, Brno, Czech Republic