Subject: Re: Removing I386_CPU from GENERIC and INSTALL
To: Thor Lancelot Simon <tls@rek.tjls.com>
From: Frank van der Linden <fvdl@netbsd.org>
List: port-i386
Date: 10/31/2006 15:16:17
On Sun, Oct 29, 2006 at 02:03:43PM -0500, Thor Lancelot Simon wrote:
> We ship other kernels which are suitable for small machines.  I propose to
> remove I386_CPU from the GENERIC and INSTALL kernels.
> 
> The I386_CPU option makes kernels considerably larger, and the VM system
> and the network stack slower.  I seriously doubt that it's possible to
> boot a GENERIC (and probably not even an INSTALL) kernel on any system
> without an i486-compatible MMU anyway.
> 
> I would also like to consider removing MATH_EMULATE from the INSTALL and
> GENERIC kernels.  It is a substantial hunk of code and, again, we have
> other kernels more suitable for small machines which include it.

Yes. For example, plain i386 CPUs haven't worked correctly with NFS for
quite a while now.  Since UBC was merged, to be specific, since the
lack of protection check on copy to/from user messes up NFS: NFS marks
pages that need to be committed (pushed out to the server) with
PG_NEEDCOMMIT, and then sets them to read-only. A plain i386 will not
notice the write fault when these pages are written to again, leading
to possible data corruption. The fact that noone had noticed this,
leads me to believe that noone uses a plain i386 anymore. I threw
away my last plain i386 last year.

So, yes to both.

- Frank