RFR JDK-8240871: SSLEngine handshake status immediately after the handshake can be NOT_HANDSHAKING rather than FINISHED with TLSv1.3

Tue May 26 21:40:42 UTC 2020

On 5/26/2020 1:26 PM, Anthony Scarpino wrote:
> On 5/13/20 1:44 PM, Xuelei Fan wrote:
>> On 5/13/2020 9:41 AM, Anthony Scarpino wrote:
>>> On 4/30/20 10:19 AM, Xuelei Fan wrote:
>>>> Hi,
>>>>
>>>> Could I get the following update reviewed:
>>>>      http://cr.openjdk.java.net/~xuelei/8240871/webrev.00/
>>>>
>>>> For TLS 1.3 full handshake, if the last handshake flight wraps the 
>>>> Finished together with other handshake message, for example client 
>>>> certificate, the flight could be wrapped and encrypted in one record 
>>>> and delegated tasks would be used.  There is no chance to return 
>>>> FINISHED handshake status with SSLEngine.(un)wrap(). However, per 
>>>> the HandshakeStatus.FINISHED specification, this handshake status is 
>>>> only generated by a call to SSLEngine.wrap()/unwrap() and it is 
>>>> never generated by SSLEngine.getHandshakeStatus().
>>>>
>>>> In order to workaround this case for TLS 1.3, the FINISHED status 
>>>> could present with SSLEngine.wrap() while delivering of the 
>>>> NewSessionTicket post-handshake message.  If this post-handshake 
>>>> message is not needed, a follow-on SSLEngine.wrap() should be called 
>>>> to indicate the FINISHED handshake status.  Although this special 
>>>> SSLEngine.wrap() should not consume or produce any application or 
>>>> network data.
>>>>
>>>> I also clean up some debug log, names and code style a little bit.
>>>>
>>>> The update could be confirmed with Tomcat and Firefox in private 
>>>> mode, as described in the bug description.  As this case happens 
>>>> only when psk_key_exchange_modes does not present, which is not a 
>>>> behavior supported by JDK, I did not find a workaround for a new 
>>>> regression test yet.  I added the labels, noreg-external and 
>>>> noreg-hard.
>>>>
>>>> Thanks,
>>>> Xuelei
>>>
>>> I do not fully understand the situation, mostly because of SSLEngine 
>>> semantics.  In normal operation, does is HandshakeStatus.FINISHED 
>>> returned when Finished is received or after the NewSessionTicket 
>>> message?
>> Not exactly.  For TLS 1.2, FINISHED will be returned with unwrap() of 
>> the finished handshake message.  However, for TLS 1.3, FINISHED will 
>> be returned any longer, because the finished handshake message is 
>> wrapped with certificate message in one record.
>>
>> For TLS 1.3:
>> 1. client send certificate, certificate verify and finished handshake 
>> message in one record.
>> 2. server call unwrap(), and return NEED_TASK to handle the 
>> certificate and certificate verify.
>>
>> So, no more FINISHED for the unwrap() return.
>>
>> It is fine if there is a after NewSessionTicket message.  The wrap() 
>> for the post-handshake message will return FINISHED.
>>
>> The bug reported is a special one that the Firefox is run in private 
>> mode, which does not request NewSessionTicket.  So there is no 
>> post-handshake generated and sent in server side.  Then, there is no 
>> FINISHED can be used if applications depends on it.
>>
>> To workaround the case, a dummy wrap() or unwrap() could be used to 
>> get the FINISHED.  The wrap() or unwrap() actually do nothing but 
>> return the FINISHED status.
>>
> 
> I don't want to be problematic, but I don't really agree with creating 
> dummy messages to generate wrap/unwrap operations in the TLS code.
No dummy message created.  Only need to call wrap() or unwrap(), but not 
data consumed or generated by the call to wrap() or unwrap(), no 
application data consumed, no network data consumed, no application data 
generated, no network data generated.

> If 
> SSLEngine is doing something wrong with not fully reading the buffer, 
> then I feel it's SSLEngine that should be fixed to handle the situation 
> right.
It is not caused by SSLEngine that does not fully reading the buffer. 
Let me try again about what's the problem.

The client (Firefox) sends, Certificate and CerticateVerify and Finished 
handshake messages in one record.  The record is encrypted.
1. One call to SSLEngine.unrwap() will read the record, and decrypt the 
record.
2. One call to SSLEngine.unwrap() cannot read the Certificate and 
CerticateVerify handshake message only, without reading the Finished 
handshake message.  It means the unwrap() method will consume the record 
data fully for all three handshake messages.
3. The Certificate and CerticateVerify should be handled in delegated 
tasks, so the call to SSLEngine.unwrap() return NEED_TASK.
4. As the SSLEngine.unwrap() return NEED_TASK, it cannot return the 
FINISHED status at the same time.
5. The FINISHED status is only be able to return with SSLEngine.unwrap() 
or SSLEngine.wrap(), and cannot returned from delegated tasks.  So the 
SSLEngine.getHandshakeStatus() after the tasks cannot be used to 
indicate the FINISHED status.
6. Then FINISHED status is missed, applications like Tomcat run into 
problem, like the bug described.

That's the problem of the bug as far as I can see.  I agreed it is not 
good to have an additional wrap() or unwrap() that did nothing, but I 
have no better idea.  It would be nice if we could have a fix in JDK 15, 
considering the impact on Tomcat and Firefox.  I'm open if there is 
alternative solution or workaround.

> Maybe not put the finished message or put all these messages 
> together.
> 
> Maybe Brad may know of a way out of this problem?  If creating a dummy 
> message is the only way to fix this, then I'm ok with it.  It is just 
> not a clean fix in my mind.
> 
> 
>>> My understanding would have been after Finished because NST is 
>>> suppose to be a post handshake message.  So in this case there is no 
>>> problem, correct?
>>>
>> Correct.
>>
>>> I'm trying to figure out why you need an empty NST.  Is the problem 
>>> when a number of messages are bundled together.  For example, the 
>>> Finished message with a partial NST, then Finished isn't processed 
>>> and both sides are waiting?  Or do both sides continue normal 
>>> traffic, it's jut the HandshakeStatus.FINISHED is one operation behind?
>>>
>> It should be fine as empty NST is just a signal to indicate the next 
>> call to wrap().  The next call to wrap() just use the signal for the 
>> return of FINISHED status, not network data produced, delivered or 
>> consumed.
>>
>>>
>>> One code comment so far:
>>> 433:  The debug message purpose was to say the NST is a stateless 
>>> ticket and not a preshared key.  Can we keep "stateless" in the message?
>>>
>> NewSessionTicket.java?  Sure, I may just want to shrink to one line.  
>> It was not intended.
> 
> shrinking the message is fine, it just needs to be clear which message 
> ticket type got sent.
> 
Added back.

Xuelei