Re: [RFC] dmaengine: add dma_slave_get_caps api

[Lars-Peter Clausen <lars@metafoo.de>]

On 07/08/2013 03:40 PM, Vinod Koul wrote:
> On Mon, Jul 08, 2013 at 02:01:35PM +0200, Lars-Peter Clausen wrote:
>> On 07/08/2013 10:54 AM, Vinod Koul wrote:
>>> +/* struct dma_slave_caps - expose capabilities of a slave channel only
>>> + *
>>> + * @src_addr_widths: bit mask of src addr widths the channel supports
>>> + * @dstn_addr_widths: bit mask of dstn addr widths the channel supports
>>> + * @src_burst_lengths: bit mask of src slave burst lengths supported
>>> + * @dstn_burst_lengths: bit mask of dstn slave burst lengths supported
>>
>> I'm not sure if we need the burst_lengths fields. For one we can only
>> express a max burst length up 32. And usually it is fine if the dma
>> controller does not support the burst length requested by the slave driver,
>> since this only specifies the maximum and the dma controller driver can
>> choose a value below this limit. E.g. if the max burst length is set to 16
>> it is still fine if the controller only supports a burst length of 8.
> well how are you picking up which one to use?
> The idea is that you query and match that with system and client to get best
> throughput. If you have IP block which over generations change capablities you
> can runtime query and then program the channel smartly!

The client would always request the largest burst size it can support, and
the dma master would then select a burst size equal or smaller to it
depending on what it can support.

E.g. if the client has an internal FIFO with room for 16 samples and it
sends a request once the FIFO is half empty, max_burst would be set to 8. If
the master supports burst sizes of 8 sample it would send 8 samples, if it
only supported burst sizes of 4 it might send either one burst of 4 or two of 4.

- Lars