Some users have reported that in polled mode the driver fails randomly
to read the last word of the transfer.

The end condition used for the transmissions (in polled and irq mode)
has been the TX_EMPTY flag. But Lars-Peter Clausen has identified a delay
from the TX_EMPTY to the actual end of the data rx.

I believe that this race condition has not been detected until now
because of the latency added by the IRQ handler or the PCIe bridge.
This bugs affects setups with low latency access to the spi core.

This patch replaces the readout logic:

For all the words, except the last one, the TX_EMPTY flag is used (and
cached).

If !TX_EMPY or is the last word. The status register is read and the
RX_EMPTY flag is used.

The performance is not affected: there is an extra read of the
Status Register, but the readout can start as soon as there is a word
in the buffer.
Reported-by: Edward Kigwana <ekigwana@scires.com>
Initial-fix-by: Lars-Peter Clausen <lars@metafoo.de>
Signed-off-by: Ricardo Ribalda Delgado <ricardo.ribalda@gmail.com>
Signed-off-by: Mark Brown <broonie@kernel.org>
Cc: stable@vger.kernel.org