Subject: Re: kernel ip_randomid() and libc randomid(3) still "broken"
To: Dennis Ferguson <dennis@juniper.net>
From: Robert Elz <kre@munnari.OZ.AU>
List: tech-net
Date: 11/27/2003 18:06:17
    Date:        Wed, 26 Nov 2003 13:51:32 -0800
    From:        Dennis Ferguson <dennis@juniper.net>
    Message-ID:  <200311262151.hAQLpXX26429@merlot.juniper.net>

  | If this is truly "*not acceptable*" then we're doomed even with the
  | full 16-bit ID space.

Yes, you're absolutely right - IP fragmentation simply has to be
eliminated (which v6 kind of attempted to do, but didn't quite
manage - though it does at least extend the ID space to 32 bits,
and only consume IDs on packets that are known to be fragmented,
not on every packet, just in case).

For UDP type applications (like NFS) that want to send big packets,
a solution is for them to use application level fragmentation,
(if simply sending smaller data units is out of the question)
where more rational approaches to reuniting the fragments, than
just a single ID, can be adopted (exactly what will depend upon
the application - but at the application level also means after
packet authentication, rather than before).

kre