Re: [PATCH v2 07/14] arm: common: edma: API to request non default queue for a channel

[Sekhar Nori <nsekhar@ti.com>]

On Tuesday 01 April 2014 06:36 PM, Peter Ujfalusi wrote:
> When using eDMA3 via dmaengine all dma channels will use the default queue.
> Since during request time we do not have means to change this it need to be done
> later, before the DMA has been started.
> With the added function it is possible to move the channel to a non default
> queue if it is possible, otherwise (when only one EQ/TC is available for the CC)
> the default queue is going to be used.
> For example: For optimal system performance the audio (cyclic) DMA should
> be placed to a separate queue which is different than what the rest of the
> system is using.
> 
> Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com>
> ---
>  arch/arm/common/edma.c             | 27 +++++++++++++++++++++++++++
>  include/linux/platform_data/edma.h |  2 ++
>  2 files changed, 29 insertions(+)
> 
> diff --git a/arch/arm/common/edma.c b/arch/arm/common/edma.c
> index be267b2080be..eaf6dd19f082 100644
> --- a/arch/arm/common/edma.c
> +++ b/arch/arm/common/edma.c
> @@ -712,6 +712,33 @@ int edma_alloc_channel(int channel,
>  }
>  EXPORT_SYMBOL(edma_alloc_channel);
>  
> +/**
> + * edma_request_non_default_queue - try to map the channel to non default queue
> + * @channel: dma channel returned from edma_alloc_channel()
> + *
> + * For certain type of applications like audio it is preferred to not use the
> + * default event queue/tc to avoid eDMA caused latency.
> + *
> + * This function will iterate through the event queues available for the CC and
> + * picks the first EQ/TC which is not set as the default for the CC
> + */
> +void edma_request_non_default_queue(int channel)
> +{
> +	unsigned ctlr = EDMA_CTLR(channel);
> +	enum dma_event_q eventq_no = EVENTQ_DEFAULT;
> +	int i;
> +
> +	for (i = 0; i < edma_cc[ctlr]->num_tc; i++) {
> +		if (i != edma_cc[ctlr]->default_queue) {
> +			eventq_no = i;
> +			break;
> +		}
> +	}
> +
> +	channel = EDMA_CHAN_SLOT(channel);
> +	map_dmach_queue(ctlr, channel, eventq_no);
> +}
> +EXPORT_SYMBOL(edma_request_non_default_queue);

With this API, the caller cannot be sure of what the final affect on the
channel will be. It depends a lot on what the default queue setting is.
And that can change independent of this API.

I am not an expert in dmaengine, but it looks like we are trying to
workaround the fact that dmaengine does not support multiple transfer
priorities for a dma_chan. If yes, it will be better to add support for
doing this in dmaengine core itself. A generic priority space could be
supported which each DMA device can then map to its hardware support.

Thanks,
Sekhar