Re: Intended usage of the dmaengine

[Dan Williams <dan.j.williams@intel.com>]

On Fri, Jul 17, 2009 at 3:33 AM, Per Forlin<per.lkml@gmail.com> wrote:
> Hi,
>
> I work with Linux in embedded systems and I got more interested in
> dmaengine after Haavard Skinnemoen added the DMA SLAVE support. DMA
> SLAVE opens up the dmaengine for embedded users and I wonder if there
> is more room for support towards the embedded world. I've made a test
> implementation of the dmaengine for our DMAC. Some parts that weren't
> supported by the dmaengine had to be exported by the driver code. A
> similar example of this is cyclic DMA jobs in dw_dmac.h.
> Is this the preferable way to handle it? Or could this functionality
> be added to the dmaengine instead?
>
> Additional support that I would like to have in the "struct dma_engine" is
> * stopping the dma channel transfer
> * continue a stopped transfer
> * PER2PER transfers (dma transfer between two peripherals)
> * dma transfers from phy mem to per, and per to phy mem
> * function to return the transfer count of an active dma transfer
> (useful when the dma channel has been stopped deliberately)
>
> I am willing to propose and contribute updates to the dmaengine
> regarding this matter. With this email I would like to check with you
> whether these types of new support are welcome in the dmaengine.

I had a similar conversation with Ira Snyder recently as his DMA_SLAVE
implementation required architecture specific extensions.  The problem
is that once you step outside pure memory-to-memory offload the
implementation rapidly gets architecture specific and quirky very
quickly.  I am not convinced that it is worth the effort to shoehorn
functionality that is by definition architecture specific into a
generic api.  Outside of providing a channel allocation scheme and a
capability for maintaining some private context for a "slave-dma"
channel I do not see a consistent role for dmaengine to play in these
embedded usage models.  You can see that the ARM/pxa developers have
arrived at a similar conclusion and are not leveraging dmaengine for
their channel management.

So I would say keep those bits you mentioned above architecture
specific for now, if we start to see generic cross-architecture
duplication then we can think about up-leveling the implementation of
those pieces.

Regards,
Dan