Re: Packet number encryption

["Brian Trammell (IETF)" <[e-mail]>]

> On 4 Feb 2018, at 21:52, Brian Trammell (IETF) <[e-mail]> wrote:
> 
>> 
>> On 4 Feb 2018, at 18:40, Christian Huitema <[e-mail]> wrote:
>> 
>> 
>> 
>> On Feb 4, 2018, at 7:18 AM, Lubashev, Igor <[e-mail]> wrote:...
>>> 
>>> Having networks wrap and unwrap packets at their boundaries is a possibility, though it has real costs and downsides. First, while networks could setup such wrapping for UDP port 443 traffic, what about QUIC traffic using other ports?  Would networks need to wrap all UDP traffic?  Second, this does reduce MTU for end users, reducing available bandwidth.  Third, this would allow networks to only answer questions like: “Is my network the culprit of the problem?” instead of “Is the problem in my network, upstream from me, or downstream from me?”.
>> 
>> This is exactly why I believe the proper place for things like spin bits or short sequence numbers is in the IP header. That would be a really neat use of the IPv6 flow ID.
> 
> Let me say this again, more directly: the IPv6 flow ID will absolutely not work for this. It's used, when it's used, for ECMP entropy and other end-of-the-path routing things. This is equivalent to the problems with ECN on older hardware / configurations using TOS for ECMP entropy: reassign these

... and things will break in unexpected and hard to debug ways.

> 
> Measurement bits aren't flow identifiers. Overloading the latter to become the former because we don't want to deal with the protocol stack we've inherited is not a serious option.
> 
> If you've got data to the contrary, please share it. But saying this will work is an exceptional claim, and requires exceptional evidence.
> 
>>> 
>>> Would it be considered outrageous to suggest that packet number encryption be an option negotiated during connection handshake?  An end user client in “Privacy-Enhanced” mode would negotiate this, while end users not particularly concerned with connection migration linkability issues would leave this off?
>> 
>> Negotiate privacy off in the name of network performance? Very slippery slope! For example, the connections that do require privacy would stick out.
>> 
>> -- Christian Huitema