Alasdair, et. al.,

Actually, I was not aware that a device could exist without a table. 
However, I suppose that geometry is a property of a device, not its 
underlying configuration, so the forced_geometry is better off in struct 
mapped_device.

Here's the third revision, with the geometry pushed into mapped_device 
as well as fixes for the problems that you pointed out wrt string 
passing, lack of warning messages, etc.  Thanks for all the great feedback!

Also, the v2 patch should have had the appropriate entries in 
include/linux/compat_ioctl.h.  Maybe it fell off?  In any case, it is 
present in this v3.

--D