Stress testing the driver with multiple start/stop events causes
kirkwood-dma to report underrun errors (which used to cause the kernel
to lock up solidly).  This is because kirkwood-i2s is not respecting
the restrictions imposed on clearing the 'pause' bit.  Follow what the
spec says; the busy bit must be read as being clear twice before the
pause bit can be released.  This solves the underruns.

However, it has been noticed that the busy bit occasionally does not
clear itself, hence the waiting is bounded to 5ms maximum to avoid a
new reason for the kernel to lockup.
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
Signed-off-by: Mark Brown <broonie@opensource.wolfsonmicro.com>