Having the SDMA driver use a tasklet for running the clients
callback introduce some issues:
  - probability to have desynchronized data because of the
    race condition created since the DMA transaction status
    is retrieved only when the callback is executed, leaving
    plenty of time for transaction status to get altered.
  - inter-transfer latency which can leave channels idle.

Move the callback execution, for cyclic channels, to SDMA
interrupt (as advised in `Documentation/dmaengine/provider.txt`)
to (a)reduce the inter-transfer latency and (b) eliminate the
race condition possibility where DMA transaction status might
be changed by the time is read.

The responsibility of the SDMA interrupt latency
is moved to the SDMA clients which case by case should defer
the work to bottom-halves when needed.
Tested-by: Peter Senna Tschudin <peter.senna@collabora.com>
Acked-by: Peter Senna Tschudin <peter.senna@collabora.com>
Reviewed-by: Vinod Koul <vinod.koul@intel.com>
Signed-off-by: Nandor Han <nandor.han@ge.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>