Re: Go Back to Single Packet Number Space

[Mikkel Fahnøe Jørgensen <[e-mail]>]

I am not currently working on handshake so I don’t have detailed insights
but I see the following concerns:
As Nick writes:


   - Each metadata object has 2 bits for tracking encryption level used to
   send the packet.


This is a problem it you create a highly optimised data structure for the
general receive case. You double or triple the state you need in bitmaps
etc. - yes the counter argument is that you need more state for
retransmission, but these indices use critical cache buffers and
retransmission is rare and can be on disk or main memory.

By having separate packet number spaces you can afford a slower less
optimized logic isolated to the handshake and simplify the logic in the
main case.

So a single number space may be simpler for simpler implementation but more
complex for an advanced implementation.

A counter argument:

The initial handshake requires more state with multiple spaces. This means
that handshake must be done a  specially allocated control block otherwise
the main connections control block will be larger than necessary. This
could be a trade-off between indirection and minimum space required for
long living silent connections.

I believe the extra initial space is not actually a problem because you can
hand-over a read connection to the main connection logic. The additional
memory required during the connection can be a concern, but it isn’t likely
a problem because:  If you are a constrained device you only deal with one
connction at a time and refuse concurrent handshakes. If you are a massive
server, you can easily handle the overhead of 100K concurrent connections
each tracking 4 packet numbers for the duration of a handshake - as long as
the overhead does not spill into the long-term connection.

And finally, keeping the handshake logic out of the main connection makes
it much simpler to experiment with alternative handshakes, for example a
simple pre-shared static key for exchanging traffic keys.

Therefore, I prefer keeping the handshake logic separate from the main
connection as much as possible.



Kind Regards,
Mikkel Fahnøe Jørgensen


On 26 July 2018 at 08.42.52, Eggert, Lars ([e-mail]) wrote:


   - Each metadata object has 2 bits for tracking encryption level used to
   send the packet.