On Wed, Jun 25, 2014 at 01:00:33PM +0100, Russell King - ARM Linux wrote: > I received a report this morning from one of the Novena developers that > the behaviour of the iMX6 ASoC codec driver (using imx-pcm-dma.c) was > sub-optimal under high system load. > > While there are issues relating to system load remaining, upon reviewing > the ASoC imx-pcm-dma.c driver, it was noticed that it not using the > residue support, because SDMA doesn't support it. This has the effect > that SDMA has to make multiple calls into the ASoC and ALSA code, one > for each period. > > Since ALSA's snd_pcm_elapsed() does not need to be called multiple times > and it is entirely sufficient to call it once to update ALSA with the > current buffer position via the pointer method, we can do better here. > We can also avoid stopping the DMA entirely, just like real cyclic DMA > implementations behave. While this means that we replay some old samples, > this is a nicer behaviour than having audio stop and restart. > > The changes to achieve this are relatively minor - imx-sdma.c can track > where the DMA is to the nearest descriptor boundary - it does this > already when deciding how many callbacks to issue. In doing this, > buf_tail always points at the descriptor which will complete next. > > The residue is defined by the bytes remaining to the end of the buffer, > when the buffer is viewed as a single block of memory [start...end]. > So, when we start out, there's a full buffer worth of residue, and this > counts down as we approach the end of the buffer, eventually becoming > zero at the end, before returning to the full buffer worth when we > wrap back to the start. > > Moving the walking of the descriptors into the interrupt handler means > that we can update the BD_DONE flag at interrupt time, thus avoiding > a delayed tasklet stopping the cyclic DMA. > > This means that the residue can be calculated from (total descriptors - > buf_tail) * descriptor size. This is what the change below does. We > update imx-pcm-dma.c to remove the NO_RESIDUE flag since we now provide > the residue. > > Signed-off-by: Russell King <rmk+kernel@xxxxxxxxxxxxxxxx> For all the audio test cases I usually run, nothing breaks, so Tested-by: Shawn Guo <shawn.guo@xxxxxxxxxx> -- To unsubscribe from this list: send the line "unsubscribe dmaengine" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html