Re: [TLS] Wrapping up cached info

[Michael D'Errico <[e-mail]>]

Stefan Santesson wrote:
> Not sure how I should interpret the silence?
> 
> I guess it means that everyone agrees with me :)

Then by that logic everyone also agrees with me to use http caching
as the model!

Seriously, I think that the current draft might be too simple.  For
example, consider caching the list of certificate_authorities in a
CertificateRequest message.  A web application might have several
different lists of certificate_authorities based on what the resource
is and who is trying to access it.  If cached-info can only cache one
object of any particular type, then it will not be very helpful for
such an application.

Also, exactly how would you hash the Diffie-Hellman parameters in a
ServerKeyExchange message?  You can't just hash the message since it
also contains a unique Ys which changes frequently if not every
handshake.

Allowing the server to instead name the objects, and keep track of
timestamps for each object, allows you to handle these two cases.  In
the multiple-CA-list example above, the server might have USER and
ADMIN lists.  If it can't be sure when exactly those lists were
created, the timestamp can simply be when the server was started or
when it last read its configuration.

When the server forces the client to renegotiate when accessing a
protected resource as an ADMIN for example, the server's cached info
extension would contain:

     type=cert_req_ca_list,name=ADMIN,timestamp=20100519123456789

When requesting a less-sensitive, but not public resource, the server
would send:

     type=cert_req_ca_list,name=USER,timestamp=20100519123456789

If the client's cached-info extension contained an exact match, then
the server would omit that data from the handshake.

This is sort of like cipher suite selection.  The client sends a list
of possible matches and the server selects one from the list.  It
differs in that there might be no match, in which case the server just
sends the info for the object it decided to send in the handshake.

I'm sure there are more details to be worked out, but I think it would
be a good idea to entertain this approach.

Mike



> To repeat in short. Here is what I suggest based on the latest traffic:
> 
> 1) A client, caching information also caches what hash algorithm was used to
> calculate the finished message at the time of caching.
> 
> 2) On repeated connections, the client indicate cached info by sending a
> hash of the cached object using the cached hash algorithm. (No use of FNV
> hash)
> 
> 3) The server accepts by returning the received hash instead of the cached
> data.
> 
> 4) The only hash agility provided is that the client will send a hash
> algorithm identifier with the hash.
> 
> 5) The client MUST NOT send more than one hash per cached object, and MUST
> used the cached hash algorithm.
> 
> 
> This solves all issues raised (securely binds the cached data to the
> finished calculation) and removes the need for hash agility
> Syntactically it just requires adding a hash identifier and adjusting the
> vector length for hash data.
> 
> So I basically wander:
> - Would this be acceptable?
> - Who could NOT live with this solution?
> - Who think it is worth the effort to agree on a better solution, and why?
> 
> Regards
>  
> 
> 
> 
> On 10-05-18 12:24 AM, "Stefan Santesson" <[e-mail]> wrote:
> 
>> As Martin pointed out to me privately, the hash used in the finished
>> calculation becomes known to the client only after receiving the serer
>> hello.
>>
>> It would therefore be natural for the client to use the hash function used
>> to calculate the finished message at the time when the data was cached.
>>
>> The client would then indicate which hash algorithm it used and upon
>> acceptance, the server will honor the request.
>>
>> /Stefan
>>
>>
>> On 10-05-17 10:34 PM, "Stefan Santesson" <[e-mail]> wrote:
>>
>>> Guys,
>>>
>>> Where I come from we have a say "don't cross the river to get to the water".
>>> And to me this proposal to change the finished calculation is just that.
>>>
>>> Look at it.
>>>
>>> The proposal is to bind the cached data by adding a hash of the cached data
>>> to the finished calculation.
>>>
>>> The proposal is further to avoid hash agility by picking the hash algorithm
>>> used by TLS's Finished message computation.
>>>
>>>
>>> Now there are two ways to achieve this goal.
>>>
>>> 1) The crossing the river to get water approach:  Exchange FNV hashes of the
>>> cached data in the handshake protocol exchange and then inject hashes of the
>>> same data into the finished calculation through an alteration of the TLS
>>> protocol.
>>>
>>>
>>> 2) The simple approach: Use the hash algorithm of the finished calculation
>>> to hash the cached data (according to the current draft).
>>>
>>>
>>> Alternative 2 securely bind the hashed data into the finished message
>>> calculation without altering the algorithm.
>>>
>>> Alternative 2 requires at most a hash algorithm identifier added to the
>>> protocol, if at all. We don't need to add negotiation since we always use
>>> the hash of the finished message calculation. Adding this identifier would
>>> be the only change made to the current draft.
>>>
>>> Alternative 2 don't require additional security analysis. If the hash used
>>> to calculate the finished message is broken, then we are screwed anyway.
>>>
>>> /Stefan
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>> On 10-05-17 9:16 PM, "Martin Rex" <[e-mail]> wrote:
>>>
>>>> Joseph Salowey wrote:
>>>>> I agree with Uri, that if you determine you need SHA-256 then you should
>>>>> plan for hash agility.  TLS 1.2 plans for hash agility.
>>>>>
>>>>> What about Nico's proposal where a checksum is used to identify the
>>>>> cached data and the actual handshake contains the actual data hashed
>>>>> with the algorithm used in the PRF negotiated with the cipher suite?
>>>>>
>>>>> This way we don't have to introduce hash agility into the extension, but
>>>>> we have cryptographic hash agility where it matters in the Finished
>>>>> computation.  Does it solve the problem?
>>>> Yes, I think so.
>>>> This approach should solve the issue at the technical level.
>>>>
>>>> Going more into detail, one would hash/mac only the data that got
>>>> actually replaced in the handshake, each prefixed by a (locally computed)
>>>> length field.
>>>>
>>>> -Martin
>>>> _______________________________________________
>>>> TLS mailing list
>>>> [e-mail]
>>>> https://www.ietf.org/mailman/listinfo/tls
>>>
>>> _______________________________________________
>>> TLS mailing list
>>> [e-mail]
>>> https://www.ietf.org/mailman/listinfo/tls
>>
> 
> 
> _______________________________________________
> TLS mailing list
> [e-mail]
> https://www.ietf.org/mailman/listinfo/tls
>