On Mon, 25 Aug 2003, Jun-ichiro itojun Hagino wrote:

> > >	- kame code shares all sys/netinet6 among all operating systems.
> > >	  therefore, if we make changes specific to netbsd (like hashes and
> > >	  CIRCLEQ) it will result in #ifdef hell.

I agree that'd be a pain.

Is there some way to abstract away the pain? I think all the different OSs
could profit from using hashes.

> > Yikes! But on the other hand, not having a hash-table and resorting to
> > a linear scan on each and every pcb lookup, means that KAME v6 is (no
> > slur intended, here) rather a bit of toy-system?  At least by
> > contemporary standards: one could not really deploy any IPv6 without
> > hashtables in the heavy-duty environments where *BSD IP/TCP is
> > currently deployed with hundreds or thousands of simultaneous [ipv4]
> > connections, at least unless you are willing to acecpt lower
> > performance levels.  Is that a fair comment?
> >
> > NB, I'm not complaining: I'm just asking. It does seem a clear
> > conseqence of comments you've made here, both recently and over the
> > past few years, consequences which had not really sunk in yet (at
> > least not for me, assuming too that my inference is valid.).
>
> 	then all 4.4BSD-based systems (i.e. without hash lookup) are toy-system.
> 	i don't think hashed inpcb lookup alone can classify operating system
> 	between toy-system and non-toy-system.  i don't see why you bother
> 	to make a comment like this.

I agree that hashed inpcb lookup alone can't qualify an OS as toy or not
toy, but if you want to make your OS perform with the best of them, you'll
need all the efficiencies you can get.

Take care,

Bill