Hi Eric,

On 02/08/15 19:52, Eric Rescorla wrote:
> I'm sure there are a few other things people would like nailed down,
> but I think the big issue here is whether or not we would require TLS 1.3
> or not.
> I would argue for not, but i can understand why people would feel the other
> way. 

So I don't get the argument for allowing any earlier versions of TLS
here at all. Why do you prefer that?

My reasons for assuming your profile, when documented, (which it is not
currently, regardless of one's familiarity with TLS) would only be a
profile of TLS 1.3 were:

- If you allow earlier versions, you get all the pain of backwards
compatibility, the need to ensure you don't have any of the old bugs,
etc. etc.

- Implementations will have to be able to respond sensibly to all sorts
of old crap ClientHello stuff at least, and maybe more. That will make
the TCPINC profile much more complex than needed, or else, make the
TCPINC profile very likely to be incompletely specified.

- I think it is also likely to increase the chances of failing to
successfully negotiate a session, simply on the basis of the number
of twistable knobs in the various versions of TLS. For example, will
there be a need for the TLS h/w accelerator bug-fixing oversized
ClientHello nonsense bytes in some cases if TCPINC goes this route?

- With earlier versions you lose whatever security analysis will be
done for TLS1.3 which is being designed to make that kind of analysis
much easier. Earlier versions are all hard to analyse, and nobody
has done work on, or afaik plans work on, analysis of this profile
over older various versions of TLS.

- Assuming QUIC is standardised and as planned also uses a profile of
TLS1.3 (albeit likely less constrained than for TCPINC), we've diverged
quite oddly.

- I assumed you'd want some 0-RTT stuff, much as I'm not a fan of
that.

- I don't see any need for almost 400 ciphersuites and all the crap
that goes with those, the TLS1.3 MTI ciphersuites should be more than
enough for TCPINC. (Too much in fact, IMO.)

- In summary, inheriting two decades of cruft for, I think, zero real
benefit, seems to me like a bad plan. There is also a major benefit in
not having every WG participant having to convince themselves that the
spec has successfully avoided that two decades of cruft via clever
profiling.

Actually the only benefit I think one can claim from allows earlier
versions of TLS would be that it makes it easier for folks who currently
have TLS code and want to re-use that for TCPINC. If that is your
argument than a) I don't find it very compelling and b) I think it means
you also need to document how it's relatively easy to take a current
generic TLS implementation and make it conform to the TCPINC profile.
(By "document" here I don't mean in an I-D though, that could be done
just fine via code or emails arguing the case.)

So - why prefer inheriting cruft? (Apologies in advance for the very
loaded question:-)

Cheers,
S.



_______________________________________________
Tcpinc mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/tcpinc

Reply via email to