[RFC] Using the DMAEngine API for scatter/gather operations

[Ira Snyder]

On Thu, May 14, 2009 at 09:27:17AM -0700, Dan Williams wrote:
> On Thu, 2009-05-14 at 09:13 -0700, Ira Snyder wrote:
> > Hello all.
> > 
> > I'm working on a driver for a custom board (based heavily on the
> > MPC8349EMDS board) to dump data out of onboard FPGA's. I need to use the
> > onboard DMA controller for this, mostly due to data transfer rate and
> > timing requirements.
> > 
> > Of course, I'd like to "play nice" with the existing Linux
> > infrastructure in my driver, even though I have no plans to submit the
> > driver upstream (since it is only useful for our custom board).
> > 
> > I have been using the DMAEngine API to interact with the DMA controller,
> > and it works without any problems. However it seems that it is missing
> > one of the major features that is supported by almost all DMA
> > controllers: scatter/gather support.
> > 
> > What I'd really like to do is give the controller a scatterlist and list
> > of hardware addresses, and have it set up the transfer. The Freescale
> > hardware can handle this, and I could program it by hand to do so, but I
> > don't want to invent my own DMA API.
> > 
> > I'm aware that the DMAEngine API was designed to offload memcpy, and
> > pretty much nothing else. Is there any chance that such a change would
> > be accepted into the mainline kernel? Any opinions on what such an API
> > should look like?
> > 
> 
> Have you taken a look at the 'slave' interface implemented by dw_dmac
> and other drivers?  It has the following interface:
> 
>        struct dma_async_tx_descriptor *(*device_prep_slave_sg)(
>                 struct dma_chan *chan, struct scatterlist *sgl,
>                 unsigned int sg_len, enum dma_data_direction direction,
>                 unsigned long flags);
> 
> Haavard uses this interface to do sgl transfers on behalf of the
> atmel-mci driver (drivers/mmc/host/atmel-mci.c).
> 

I had started looking at it, though the Freescale device does not
implement DMA_SLAVE support. I was a bit discouraged by the changelog
for commit dc0ee643, which says that the interface is for "DMA with
hardware handshaking", which I do not have. There is not anything
special about accessing the FPGA's, they appear in the processor's
memory map and can be read from/written to just like RAM.

I'll try studying the dw_dmac implementation and atmel-mci example.

Thanks for the response,
Ira