Re: New Version Notification for draft-bonica-6man-frag-deprecate-00.txt

[Kevin Lahey <[e-mail]>]

On Tue, 25 Jun 2013 16:36:51 +0000
Christian Huitema <[e-mail]> wrote:

> > I also believe that FreeBSD has done the best it can, and reasonably so. It is debatable whether a ICMP6 PTB message should apply to all currently open TCP sessions to the same destination, as I wonder about multi-path TCP and path diversity here.
> 
> There are certainly different ways to implement that, without inserting a fragmentation header. For example, if the lower layer knows that the packet is too big, it could send the local equivalent of "ICMP too big" back to the TCP layer.

This certainly seems like a bug, no matter how you slice it:

* If the implementation wanted to allow for different MTUs on a per-flow
  basis (reasonable given the prevalence of NATs), it should go ahead 
  and send a full-sized packet for the second flow, and find out what
  the MTU is.

* If the implementation wanted to efficiently use the first Packet Too
  Big it got to fix all possible flows, TCP should re-segment the data
  in the second flow, rather than allowing fragmentation.

I can't think of a case in which it would be better for TCP to fragment
rather than re-segment (since it's a lot more efficient to recover from
a dropped segment than a dropped fragment), but I'm sure I will as soon
as I hit send.

Kevin
[e-mail]