>Risk is actually fairly small.  Figure:
>
>1.  The system is in single-user mode, so nothing is actively writing.
>2.  We have sync'd.
>3.  We then write to the raw (uncached) disk device.
>4.  After a short delay, when the disk light is done flashing, we depower the
>system.

Agreed, risk is low, but here, isn't the Right Thing really to allow
the kernel to update its incore superblock and write that to disk?