RFR 8211018: Session Resumption without Server-Side State
Xuelei Fan
xuelei.fan at oracle.com
Thu Jun 6 13:09:34 UTC 2019
On 6/5/2019 11:04 PM, Jamil Nimeh wrote:
> Hi Xuelei, thanks for the feedback. A couple comments in-line below.
>
> On 6/5/2019 5:37 PM, Xuelei Fan wrote:
>>
>>
>> On 6/5/2019 4:57 PM, Jamil Nimeh wrote:
>>> I think what it's saying is that what was explicitly called out in
>>> 4507 (where there is both the extension_data length bytes AND the
>>> opaque vector length) is not how deployed implementations did it. It
>>> implies that deployed implementations do what you laid out below
>>> where you just have 2 bytes of ID and 2 bytes of length. And I
>>> believe that is compatible with what 5077 specifies.
>>>
>> Hm, I agreed with you.
>>
>>> So the potential problem is if one endpoint or the other happens to
>>> implement 4507 to the letter, extra length bytes and all. But the
>>> authors of 5077 say that no known implementations do this. That's
>>> good for us I think, because in the two-ish years between 4507 and
>>> 5077 nobody did straight 4507, or they maybe did and fixed it by the
>>> time 5077 came around.
>>>
>> I may view it differently.
>>
>> If an implementation encoded per the format:
>> 00 23 Extension type 35
>> 00 02 Length of extension contents
>> 00 00 Length of ticket
>>
>> Just as your analysis previous, a RFC 5077 server will just ignore the
>> extension. No real hurt actually.
>>
>> But if an implementation encoded empty SessionTicket extension per the
>> format (the known implementation):
>> 00 23 Extension type 35
>> 00 00 Length of extension contents
>>
>> The server could read it as RFC 5077, and use stateless
>> implementation. When the ServerHello extension sent back. No matter
>> the RFC 4057 client accept it or not, there are interop issues.
>>
>> If the client does not accept it (unlikely), the connection cannot be
>> established.
> JN: Not necessarily. If the client doesn't accept it that doesn't mean
> that it has to be fatal to the handshake.
We may be not talking about the same scenarios. Let me try again.
1. a RFC 4507 client send an empty ST extension for initial handshaking
in the format:
00 23 Extension type 35
00 00 Length of extension contents
2. a REFC 5077 server response with an en empty ST extension in ServerHello.
00 23 Extension type 35
00 00 Length of extension contents
3. if the RFC 4507 client cannot accept the ST in ServerHello, instead
of ignore the extension, the client will reject with a fatal alert and
then close the connection. Even if the client choose to ignore the ST
in ServerHello, I'm not sure if it could handle the coming
NewSessionTicket handshake message.
However, #3 is unlikely to happen because the client is already use this
format for the ClientHello message for empty SH message. I think
appendix A in RFC 5077 describe that the known RFC 4507 implementation
works. If #3 does not happen, it means the client accept the DH
message, and can establish the connection.
So we can rewrite 3 as 3':
3'. the RFC 4507 client accept the ST and NST, and the connection
established.
4. The RFC 4507 client uses the following format for session resumption:
00 23 Ticket Extension type 35
01 02 Length of extension contents
01 00 Length of ticket
FF FF .. .. Actual ticket
5. The RFC 5077 server cannot parse the ST extension for the session
resumption, and the session resumption cannot be performed.
> It just means resumption
> cannot happen. I'd need to look at the spec to see what it says (if it
> covers this edge case) but just thinking about it logically, a malformed
> ticket shouldn't have to force an alert. It doesn't if we receive a
> malformed ticket (Tony recently improved the logic around that ticket
> processing to make sure of that), but it shouldn't mean the existing
> handshake cannot otherwise complete.
>>
>> If the client accept it, the resumption session will use the
>> negotiated ticket, and then non-empty SessionTicket extension is
>> encoded as:
>> 00 23 Ticket Extension type 35
>> 01 02 Length of extension contents
>> 01 00 Length of ticket
>> FF FF .. .. Actual ticket
>>
>> The server would have to handle it (RFC 4507 format) if it want the
>> session resumption works. Here come the interop issues.
> JN: Okay, I can see how this sequence of events could play out. That
> said, I think before we contort ourselves into a pretzel to accommodate
> this edge case, there should at least be some client out there that
> people actually use (not some archived dust-bunny) that we can find that
> behaves in this manner. Otherwise, my feeling is that we should just
> support session tickets per 5077 and not any older obsoleted specification.
Sorry, I did not make me clear. I did not mean to support an obsoleted
specification. I meant that we could break the existing RFC 4507 based
applications.
Without this fix in server side, the existing RFC 4507 based client
application works as expected. It can establish the initial connection,
and can perform session resumption.
However, if the server upgrade with this patch, the existing RFC 4507
based client application cannot perform session resumption any more.
That's a compatibility impact to me.
Hope it is a little bit clear. Otherwise, let's have a meeting today.
Xuelei
>>
>> Xuelei
>>
>>> I dug up a few pcaps I've kept around during testing of other TLS
>>> features over the past few years. I had Chrome, Mozilla and OpenSSL
>>> s_client pcaps and they all appear to follow the 5077 format. Of
>>> course, anything after 2008 is more likely to do 5077 than 4507.
>>>
>>> --Jamil
>>>
>>> On 6/5/2019 4:35 PM, Xuelei Fan wrote:
>>>> I'm not sure I understand the following words in page 17, RFC 5077.
>>>>
>>>> " An error in the encoding caused the specification to differ from
>>>> deployed implementations. At the time of this writing there are no
>>>> known implementations that follow the encoding specified in RFC
>>>> 4507.
>>>> "
>>>>
>>>> Is it means that the known implementation encode empty SessionTicket
>>>> extension as?
>>>> 00 23 Extension type 35
>>>> 00 00 Length of extension contents
>>>>
>>>> Xuelei
>>>>
>>>> On 6/5/2019 4:26 PM, Xuelei Fan wrote:
>>>>> On 6/5/2019 3:37 PM, Jamil Nimeh wrote:
>>>>>> I think we're overstating the "otherwise" case. A client that
>>>>>> uses this strict 4507 format would initially send a ticket that
>>>>>> looks like { 00 23 00 02 00 00 } to which our server would reject
>>>>>> this extension (since the final 00 00 would be interpreted as a
>>>>>> ticket when the client did not intend it to be so). The result of
>>>>>> this SHOULD be that the server responds with a ServerHello that
>>>>>> doesn't have the SessionTicket extension.
>>>>>>
>>>>>> That doesn't mean that resumption cannot happen. It just means
>>>>>> that resumption happens using the usual session ID lookup approach
>>>>>> and the server is caching the session rather than letting the
>>>>>> client do the work. Given that this is a degenerate case for what
>>>>>> (I hope) is a small subset of older clients, I think using
>>>>>> server-cached sessions is OK.
>>>>>>
>>>>>> I don't believe we should ever find ourselves in a case where the
>>>>>> strict-4507 client actually gets a real ticket from our server,
>>>>>> and in turn should never hand us a ticket thinking that resumption
>>>>>> could actually take place via said ticket.
>>>>>>
>>>>> I'm not very sure if I read the Appendix A of RFC 5077 correctly. I
>>>>> think it is trying to explain that client does not use strict-4507
>>>>> for the empty extension and then result in the interop issues.
>>>>>
>>>>> Page 18, RFC 5077:
>>>>> " Note that the encoding of an empty SessionTicket extension was
>>>>> ambiguous in RFC 4507. An RFC 4507 implementation may have
>>>>> encoded
>>>>> it as:
>>>>>
>>>>> 00 23 Extension type 35
>>>>> 00 02 Length of extension contents
>>>>> 00 00 Length of ticket
>>>>>
>>>>> or it may have encoded it the same way as this update:
>>>>>
>>>>> 00 23 Extension type 35
>>>>> 00 00 Length of extension contents
>>>>> "
>>>>>
>>>>>> On the client side, we cannot know ahead of time that the server
>>>>>> is strict-4057, so we have to send a 5077 style SessionTicket
>>>>>> extension. The server will probably not like that and not assert
>>>>>> SessionTicket in its server hello. So our client should fall back
>>>>>> to using the session ID to support resumption, and the server
>>>>>> should follow suit by caching the session.
>>>>>>
>>>>> I agreed. We should stick to the RFC 5077 format in client side.
>>>>>
>>>>> Thanks,
>>>>> Xuelei
>>>>>
>>>>>> --Jamil
>>>>>>
>>>>>> On 6/5/2019 2:28 PM, Xuelei Fan wrote:
>>>>>>> I don’t know if there are any deployment of RFC 4507. If not, we
>>>>>>> are safe; otherwise there are interop problems for session
>>>>>>> resumption.
>>>>>>>
>>>>>>> Xuelei
>>>>>>>
>>>>>>>> On Jun 5, 2019, at 2:19 PM, Jamil Nimeh
>>>>>>>> <jamil.j.nimeh at oracle.com> wrote:
>>>>>>>>
>>>>>>>> Hi Xuelei,
>>>>>>>>
>>>>>>>> Given that 4507 is obsoleted in favor of 5077 is there really
>>>>>>>> that much value to supporting this older/broken extension
>>>>>>>> format? Do we know of clients that still adhere to 4507?
>>>>>>>> Otherwise it seems better to stick to 5077 and the approach in
>>>>>>>> TLS 1.3 and not try to go back and support an earlier obsoleted
>>>>>>>> approach to this feature.
>>>>>>>>> These lines took me to the cooperation behaviors between RFC
>>>>>>>>> 5077 and RFC 4507. It looks like we don't support RFC 4507
>>>>>>>>> format of SessionTicket extension. As RFC 5077 and RFC 4507
>>>>>>>>> use the same extension ID for different extension format. There
>>>>>>>>> are potential compatibility issues, and make session resumption
>>>>>>>>> impossible. I would like to have a workaround to accept both
>>>>>>>>> formats. For example, using the a cookie at the beginning of
>>>>>>>>> the ticket, as described in appendix-A of RFC 5077.
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> I will review the rest of this class in the afternoon or tomorrow.
>>>>>>>>>
>>>>>>>>> Thanks,
>>>>>>>>> Xuelei
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>
>>>
More information about the security-dev
mailing list