Uncorrectable errors (ECC) with jffs2

I'm trying to get my root filesystem into the BeagleBoard's NAND
flash. I followed the recommended procedure outlined here:

http://elinux.org/BeagleBoardNAND#Writing_file_system_with_kernel

When I extract my tarball onto the partition, I get a few errors like
this:

mtd->read(0x44 bytes from 0xf9406fc) returned ECC error

However, most of the files appear to be okay (I can run my test
applications etc.).

When I do "sync" and unmount the partition and then remount it again,
every read seems to be uncorrectable; my console gets flushed with ECC
errors.

I'm at a loss here, since creating a jffs2 image with mkfs.jffs2
produced the same errors. Since there are so many things that might
have caused this, I desperately need a pointer in the right direction.

Some additional info, please ask for more if that might help:

I'm using a linux-omap3 kernel from the arago git repository and
checked out the 2.6.33.7-rt29 tag (I need a PREEMPT-RT kernel). The
rootfs is built from scratch and contains only BusyBox (1.18.4 built
from source), libc and qt. The MTD-utils used on the BeagleBoard are
the ones built into BusyBox. I'm using mdev to manage devices. My
toolchain is CodeSourcery arm-2009q3 (linux-gnueabi).

The kernel has been reconfigured many times, this would be my first
guess at what is wrong. When I started using it, MTD support wasn't
even selected, but I don't really know, which options have to be
configured for an OMAP processor. If it helps, I can provide my kernel
config...

Any help would be greatly appreciated!

Regards
Torben Frenzel

Nevermind, I found the problem:

CONFIG_MTD_NAND_OMAP_PREFETCH=y

was set and apparently, this doesn't work well with my kernel.
However, if anyone has a suggestion what I have to do to use prefetch
support to increase performance, please let me know.

Regards
Torben Frenzel