Re: Is there a binding for IORESOURCE_DMA population?

[Grant Likely <grant.likely-s3s/WqlpOiPyB63q8FvJNQ@public.gmane.org>]

On Sat, Jul 16, 2011 at 02:09:46PM +0200, Arnd Bergmann wrote:
> On Saturday 16 July 2011 09:57:49 Shawn Guo wrote:
> > If I understand this correctly, this is something I demonstrated in
> > the initial message.  I do not think we always have individual
> > device_node for each channel (struct dma_chan).  The most common
> > situation is all channels are backed by one device_node (dmaengine).
> > 
> > As I stated in the initial message, (for i.mx sdma example) the
> > problem is 'struct dma_chan' stands for a physical dma channel, while
> > all platform_device has is a virtual channel number (known as sdma
> > event) assigned by hardware.  This virtual channel can only be
> > dynamically mapped to a physical one by sdma driver.  That said there
> > is no way for us to specify a physical channel in device tree.  We can
> > only specify the virtual one there.  (I think this is kinda common,
> > and sdma is just one example.)
> 
> My assumption was that you would create a device_node for each channel
> in the device tree source, but not necessarily have a platform_device
> for each of them.
> 
> > i.mx sdma has 32 physically channels and 48 virtual ones (events).
> > If we go the way you are suggesting, we will have to have 48 nodes
> > under sdma node just for giving event number, and then have
> > dma-channels property of device specifying the phanldles to these
> > event nodes belonging to it.
> > 
> > So we get something like below, which looks silly to me.
> > 
> > sdma@83fb0000 {
> >         #address-cells = <1>;
> >         #size-cells = <0>;
> >         compatible = "fsl,imx51-sdma", "fsl,imx35-sdma";
> >         reg = <0x83fb0000 0x4000>;
> >         interrupts = <6>;
> >         fsl,sdma-ram-script-name = "sdma-imx51.bin";
> > 
> >         dma00: dma-channel@0 {
> >                 compatible = "fsl,sdma-channel";
> >                 reg = <0>;
> >         };
> > 
> >         dma01: dma-channel@1 {
> >                 compatible = "fsl,sdma-channel";
> >                 reg = <1>;
> >         };
> > 
> >         ......
> > 
> >         dma47: dma-channel@47 {
> >                 compatible = "fsl,sdma-channel";
> >                 reg = <47>;
> >         };
> > };
> > 
> > ssi@83fcc000 { /* SSI1 */
> >         compatible = "fsl,imx51-ssi", "fsl,imx1-ssi";
> >         reg = <0x83fcc000 0x4000>;
> >         interrupts = <29>;
> >         fsl,ssi-uses-dma;
> >         dma-channels = <&dma00, &dma01>;
> > };
> > 
> > Did I state the problem clear?  Or am I missing anything?
> 
> Right, this is what I had in mind. I think this is reasonable.
> If you need so many channels, the device tree will be huge already,
> so this doesn't add that much bloat either.
> 
> Another option would be to add a 'dma-parent' property akin to the
> interrupt-parent property and then just refer to the channel
> numbers within the parent dma controller.
> I guess you would prefer something like that in order to keep the
> device tree source more compact, right?

Rather than a dma-parent property, I'd recommend following approach
that the gpio binding uses and have:

	#dma-channel-cells = <n>

...in the dma controller, and...

	dma-channels = <&dma-controller (dma-specifier)>;

...in the dma user.  Then the dma controller binding can specify
how much data is required for a dma-specifier (number of cells), and
what each cell means.

g.