This allows PCI devices that can only address (e.g.) 36 or 40 bit DMA to
use direct DMA, at the cost of not being able to DMA to non-RAM addresses
(this doesn't affect MSIs as there is a separate dedicated window for
that) which we wouldn't have been able to do anyway if the RAM size didn't
trigger the creation of the second inbound window.

It also fixes an off-by-one error that set dma_direct_ops on PCI devices
whose dma mask could address all the space below the DMA offset
(previously 40 bits), but not the window that starts at the DMA offset.
Signed-off-by: Scott Wood <oss@buserror.net>
Cc: Tillmann Heidsieck <theidsieck@leenox.de>
Tested-by: Tillmann Heidsieck <theidsieck@leenox.de>