Re: [TLS] Fwd: New Version Notification for draft-kazuho-protected-sni-00.txt

[Kazuho Oku <[e-mail]>]

Hi,

Thank you for your comments.

2017-07-19 7:05 GMT+02:00 Ilari Liusvaara <[e-mail]>:
> On Wed, Jul 19, 2017 at 05:42:24AM +0200, Kazuho Oku wrote:
>> Hi,
>>
>> I am happy to see us having discussions on how to protected SNI. I am
>> also happy to see that draft-huitema-tls-sni-encryption [1] proposes
>> actual methods that we might want to use, and that the I-D discusses
>> about various attack vectors that we need to be aware of.
>>
>> On the other hand, as stated on the mailing list an on the mic, I am
>> not super happy with the fact that the proposed methods have a
>> negative impact on connection establishment time.
>>
>> So here goes my straw-man proposal, as an Internet Draft:
>> https://datatracker.ietf.org/doc/draft-kazuho-protected-sni/.
>
> At least that method does not obviously vulernable to replay. However,
> it has multitude of other problems:
>
>> In essence, the draft proposes of sending information (e.g.,
>> semi-static (EC)DH key) to bootstrap encryption in ClientHello as a
>> DNS record. Clients will use the obtained (EC)DH key to encrypt SNI.
>
> There are multiple problems with use of DNS:
>
> - Individual DNS records can be at most 64kB.
> - The whole DNS reply can be at most 64kB.
> - In practice, exceeding about 1200 bytes for response starts
>   causing problems with DNSoDTLS (there is a fallback to DNSoTLS,
>   but it is not very pleasant).

Yes. It is very unlikely that a TLS bootstrap record that contains a
signature and a certificate chain will fit into a single packet, hence
there would be a performance penalty if DNS over DTLS that fallback to
TLS is used.

I have had DNS over QUIC in mind when writing the draft, and I might
say that the draft (especially the mode that bootstraps the _signed_
signature) is written under the premise that DNS over QUIC will become
popular in the future.

> - The textual representation starts getting unpleasant way before
>   that.
> - DNS records are defined to have textual and wire formats. Using TLS
>   notation would be pretty great for latter, but not good for the
>   former.

Generally speaking, for the following two reasons, I prefer having a
TLS structure that defines how a bootstrap information should be
encoded, and then have a DNS record type that conveys the information.

First reason is because the certificate chain and the signature that
signs the EC(DH) keys ideally should be validated the same way as we
do with the server certificates transmitted within a TLS handshake. It
would make sense to just convey certain extensions that affect the
behavior of validation (e.g., OCSP stapling extension) as-is, rather
than trying to come with corresponding textual representations.

Second reason is that we might want to transmit the bootstrap record
using something other than DNS.

>> Since DNS queries can run in parallel, there would be no negative
>> performance impact, as long as DNS responses can be obtained in a
>> single RTT.
>
> Unfortunately, this is not quite true:
>
> - Packets can get lost, and DNS retransmissions are fairly slow.
> - There are lots of horked DNS servers that respond in all sorts of
>   bad ways (including timing out, or sending utterly bogus error
>   codes) to unknown RRTypes, especially if the RRType number is
>   above 255 (but bad servers that do that to any RRType they do
>   not know about still exist).

If DNS over QUIC (or TLS) is used to transfer a signed bootstrap
record with a certificate chain, the amount of data that is sent (in
sum of DNS query and TLS handshake) will be roughly the same, assuming
that Cached Information Extension (RFC7924) is used. Therefore, I
believe that the performance will be roughly the same even when taking
packet loss into consideration.

If UDP-based DNS protocol is used to transfer a signed bootstrap
record with a certificate chain, the DNS query needs to fallback to
TCP, since the record does not fit into a single packet. That means
that it would never be possible to obtain the DNS response in 1 RTT in
such case.

If UDP-based DNS protocol is used to transfer a bootstrap record
without a sign or a certificate chain, then the question will be how
high the probability will be in the following two cases: lose more
than 1 packet within 2 packets vs. lose more than 1 packet within 4
packets. I agree that amortized performance will be worse for the
latter.

> (And let's also ignore difficulties in adding new RRTypes, despite
> RFC3597 being almost 14 years old, and covering almost everything that
> has been added (one needs extra justfication for breaking RFC3597-
> complatiblity[1]).)
>
>
> [1] In practicular, following kinds of things break RFC3597:
>
> - Compressing RRDATA with reference to rest of the DNS message (e.g.,
>   DNS name compression).
> - Requiring any kinds of transofrmations on the data in path between
>   the master file and the end application.
>
>> The draft mainly discusses about sending a signed bootstrap
>> information together with the certificate chain, since doing so is not
>> only more secure but opens up other possibilities in the future (such
>> as 0-RTT full handshake). However, since transmitting a bootstrap
>> record with digital signature and identity is unlikely to fit in a
>> single packet (and therefore will have negative performance impact
>> until DNS over TLS or QUIC becomes popular), the draft also discusses
>> the possibility of sending the EC(DH) key unsigned in the "Things to
>> Consider" section.
>
> There is not much space in DNS records for anything bigger than one
> ECDH key per record (but there is space for a few records of that
> kind).

Agreed.

>
> -Ilari



-- 
Kazuho Oku