libata made the assumption that (for PIO commands in this case)
it could modify DMA memory at the kernel-virtual address, after
mapping this.  This is incorrect, and fails on e.g. platforms that
copy DMA memory back and forth (swiotlb on Intel EM64T and IA64).

Remove this assumption by ensuring that we only call the DMA mapping
routines if we really are going to use DMA for data xfer.

Also:  remove a bogus WARN_ON() in ata_sg_init_one() which caused
bug reports (but no problems).