[PATCH RFCv2 1/2] dmaengine: add support for scatterlist to scatterlist transfers

Dan Williams dan.j.williams at intel.com
Tue Sep 28 03:35:09 EST 2010

On Mon, Sep 27, 2010 at 10:23 AM, Ira W. Snyder <iws at ovro.caltech.edu> wrote:
> On Mon, Sep 27, 2010 at 05:23:34PM +0200, Linus Walleij wrote:
>> 2010/9/25 Ira W. Snyder <iws at ovro.caltech.edu>:
>> > This adds support for scatterlist to scatterlist DMA transfers.
>> This is a good idea, we have a local function to do this in DMA40 already,
>> stedma40_memcpy_sg().
> I think that having two devices that want to implement this
> functionality as part of the DMAEngine API is a good argument for making
> it available as part of the core API. I think it would be good to add
> this to struct dma_device, and add a capability (DMA_SG?) for it as
> well.
> I have looked at the stedma40_memcpy_sg() function, and I think we would
> want to extend it slightly for the generic API. Is there any good reason
> to prohibit scatterlists with different numbers of elements?
> For example:
> src scatterlist: 10 elements, each with 4K length (40K total)
> dst scatterlist: 40 elements, each with 1K length (40K total)
> The total length of both scatterlists is equal, but the number of
> scatterlist entries is different. The freescale DMA controller can
> handle this just fine.
> I'm proposing this function signature:
> struct dma_async_tx_descriptor *
> dma_memcpy_sg(struct dma_chan *chan,
>              struct scatterlist *dst_sg, unsigned int dst_nents,
>              struct scatterlist *src_sg, unsigned int src_nents,
>              unsigned long flags);
>> > This is
>> > currently hidden behind a configuration option, which will allow drivers
>> > which need this functionality to select it individually.
>> Why? Isn't it better to add this as a new capability flag
>> if you don't want to announce it? Or is the intent to save
>> memory footprint?
> Dan wanted this, probably for memory footprint. If >1 driver is using
> it,

Yes, I did not see a reason to increment the size of dmaengine.o for
everyone if only one out-of-tree user of the function existed.

> I would rather have it as part of struct dma_device along with a
> capability.

I think having this as a dma_device method makes sense now that more
than one driver would implement it, and let's drivers see the entirety
of the transaction in one call.


More information about the Linuxppc-dev mailing list