Re: DMA APIs gumble grumble

[Benjamin Herrenschmidt <benh@kernel.crashing.org>]

On Wed, 2006-11-08 at 10:25 +0200, Muli Ben-Yehuda wrote:
> On Wed, Nov 08, 2006 at 12:54:37PM +1100, Benjamin Herrenschmidt wrote:
> 
> >  - For platforms like powerpc where I can have multiple busses on
> > different IOMMU's and with possibly different DMA ops, I really need to
> > have a per-device data structure for use by the DMA ops (in fact, in my
> > case, containing the DMA ops themselves). Right now, I defined a notion
> > of "device extension" (struct device_ext) that my current implementation
> > puts in device->firmware_data (don't look for it upstream, it's all
> > patches queuing up for 2.6.20 and about to go into powerpc.git), but
> > that I'd like to have flat in struct device instead. Would it be agreed
> > that linux/device.h includes itself an asm/device.h which contains a
> > definition for a struct device_ext that is within struct device ? That
> > would also avoid a pointer indirection which is a good thing for DMA
> > operations
> 
> I want multiple dma_ops for Calgary on x86-64, so strong thumbs up for
> doing this in a generic manner. device_ext kinda sucks as a name,
> though... if it's used for just dma_ops, how about device_dma_ops?
> 
> I agree with the rest of your suggestions too, FWIW.

Yes, I need multiple dma_ops for powerpc too and additional void * data
that go with them (iommu instance).

I use it for more than dma ops though. I posted the actual structure
content in another reply to Dave.

I agree, though, device_ext sucks as a name, you are welcome to propose
something better, I'm no good at finding names :-)

Ben.