On Thu, May 19, 2005 at 11:06:11PM +0200, theo borm wrote:
> dear list members,
> 
> I've just spent 2 days trying to figure out what is going wrong with
> my diskless cluster nodes, and have become utterly frustrated.
> 
> I have a cluster of diskless nodes: cheap i386 (actually AMD XP2600's)
> with a reasonable amount of memory, onboard VGA and via rhine (vr0)
> NIC, and an extra realtek 8169S (re0) PCI NIC. I would very much like
> to be able to netboot a stable NETBSD release (production system; no
> time to chase current), *and* be able to use the gigabit connectivity.
> Currently the nodes work, with a lot of buts and ifs (NFS and vr0
> do not seem to like each other) under 1.6.2, which does not support the
> realtek re0's, so I want to migrate the nodes to a 2.x series kernel.
> The front end to this cluster is an NFS server which currently runs
> 2.0.2. It exports individual root directories to the diskless nodes.
> 
> Replacing the 1.6.2 installation in diskless cluster nodes with
> fresh 2.0.2 installations (from a vanilla 2.0.2. release distribution)
> results in the system returning "exec /sbin/init: error 8", eventually
> resulting in a kernel panic "panic: no init"
> 
> [...]
> 
> If I replace the 2.0.2. /sbin/init by the 1.6.2 /sbin/init, this file
> /is/ properly executed (COMPAT 16), after which the system
> tries to drop into single user mode, but can't because /bin/sh is
> not in the correct executable format. Only if I downgrade all
> bin & sbin binaries to 1.6.2 does the system (more or less) boot
> again.

It could be a shared library issue. In 2.0 all binaries (including /sbin/init)
are dynamically linked, while /bin and /sbin are statically linked in 1.6.x.
Make sure you have a properly-populated /lib and /usr/lib

-- 
Manuel Bouyer <bouyer@antioche.eu.org>
     NetBSD: 26 ans d'experience feront toujours la difference
--