Re: [tcpinc] Adam Roach's Yes on draft-ietf-tcpinc-tcpcrypt-09: (with COMMENT)

[Adam Roach <[e-mail]>]

On 11/17/17 11:14, Daniel B Giffin wrote:
> Adam Roach wrote:
>> On 11/9/17 08:16, Kyle Rose wrote:
>>> On Thu, Nov 9, 2017 at 10:11 AM, Adam Roach <[e-mail]> wrote:
>>>> On 11/8/17 19:45, Mirja Kuehlewind (IETF) wrote:
>>>>> That’s not true. This is to cover the case where the packet got corrupted on
>>>>> the path, thus hopefully the retransmission will decrypt correctly.
>>>> So, to be clear, you're talking about packet corruption that happens to
>>>> produce a valid checksum, right? If that's the reasoning here, the authors
>>>> probably want to include that rationale in the document.
>>> Mirja's is my interpretation, as well.
>>>
>>> Off-path attackers wouldn't be able to sent segments with the right
>>> sequence number with high probability, so it's unlikely that this is a
>>> DoS vector; but giving implementations the option of simply dropping
>>> segments for which the authenticity check fails is not likely to cause
>>> recurring timeout problems for correct implementations.
>> Yes, and you're giving implementors options here without really explaining
>> why -- which means they're probably just going to pick one randomly. Adding
>> text that gives some notion about why one might choose one option over the
>> other would allow them to make an informed choice rather than a random one.
> That makes sense.  How about the following text instead?
> (It is very similar to how we treat spurious FIN in the
> following section, 3.7.)
>
>    [...] The output of this operation is either a plaintext
>    value `P` or the special symbol FAIL.  In the latter case,
>    the implementation SHOULD abort the connection and raise
>    an error condition distinct from the end-of-file
>    condition.  But if none of the TCP segment(s) containing
>    the frame have been acknowledged and retransmission could
>    potentially result in a valid frame, an implementation MAY
>    instead drop these segments.
>
> This doesn't speculate on how the corruption came about (I
> agree it's difficult to find a reasonable case), but it
> gives a reasonable default via SHOULD and also leaves a
> little room for implementations with some sort of clever
> anti-DOS mitigation.

Thanks -- this seems like a good approach.

/a