cvs commit: src/sys/netinet ip_fastfwd.c ip_input.c ip_var.h

[Luigi Rizzo]

On the principle, I tend to agree with Darren here...
it is not nice to replicate functionality in multiple places
by using specialized code instead of relying on (and
possibly optimizing) the generic one. It makes a lot harder
to clean up the replication later, and i believe Andre knows
that quite well given the cleanup work he has done in the past
in the network stack.

I don't think it is worth making a bit fuss about this particular
change, but certainly, as a general principle, we should try as
much as possible to use the generic mechanisms when available --
especialliy given that performance killers are elsewhere (locking
etc.).

	cheers
	luigi

On Sat, May 08, 2004 at 08:25:31AM -0700, Darren Reed wrote:
> On Fri, May 07, 2004 at 07:55:36AM -0700, Sam Leffler wrote:
> > 
> > Employing a packet filter is not equivalent as it requires every packet to be 
> > processed while this (effectively 7-line change) adds no new overhead to the 
> > normal processing path for packets.  It would be nice if packet filtering 
> > were cheap enough that we could use it in this way but I don't think that's 
> > the case just yet.
> 
> Using that argument, is that clearance to put all of the normalization
> from pf into the various parts of the networking code (not every type of
> normalisation needs to be done on every packet but it is all useful), with
> sysctls to turn it on or off, and maybe we'll add the ability to log packets
> at various points because we don't want the overhead of BPF (it has to
> process every packet too) and that's just for starters.  I'm sure I can
> think of some more, in time.  How about you?
> 
> If there were a core@ for freebsd that was active, this is the kind of
> thing I'd be writing to them about, asking for it to be backed out.
> 
> Darren