DMA handles all data transfers up to 128K or 16 segments and fallback to
pio mode when DMA requirements are not met.
The read performance has improved greatly while the write performance also
showed some improvement but seems limited by the card type and quality.
Submitted by: Pratik Singhal <pratiksinghal@freebsd.org>
Sponsored by: Google Summer of Code 2015
Tested on: A10 (cubieboard) and A20 (cubieboard 2 and banana pi)
77e5d907d2