The hardware supports transfers up to a page boundary per buffer.
Currently, we work around that in the DMA code by splitting each
buffer up as we run through the scatterlist.  Avoid this by telling
the block layers about the hardware restriction.

Eventually, this will allow us to phase out the splitting code,
but not until the old IDE layer allows us to control the value it
gives to blk_queue_segment_boundary().
Signed-off-by: NRussell King <rmk+kernel@arm.linux.org.uk>