Re: [PATCH v3 4/7] of: configure the platform device dma parameters

[Arnd Bergmann <arnd@arndb.de>]

On Tuesday 27 May 2014 13:56:55 Grant Likely wrote:
> On Fri, 02 May 2014 11:58:30 +0200, Arnd Bergmann <arnd@arndb.de> wrote:
> > On Thursday 01 May 2014 14:12:10 Grant Likely wrote:
> > > > > I've got two concerns here. of_dma_get_range() retrieves only the first
> > > > > tuple from the dma-ranges property, but it is perfectly valid for
> > > > > dma-ranges to contain multiple tuples. How should we handle it if a
> > > > > device has multiple ranges it can DMA from?
> > > > > 
> > > > 
> > > > We've not found any cases in current Linux where more than one dma-ranges
> > > > would be used. Moreover, The MM (definitely for ARM) isn't supported such
> > > > cases at all (if i understand everything right).
> > > > - there are only one arm_dma_pfn_limit
> > > > - there is only one MM zone is used for ARM
> > > > - some arches like x86,mips can support 2 zones (per arch - not per device or bus)
> > > >   DMA & DMA32, but they configured once and forever per arch.
> > > 
> > > Okay. If anyone ever does implement multiple ranges then this code will
> > > need to be revisited.
> > 
> > I wonder if it's needed for platforms implementing the standard "ARM memory map" [1].
> > The document only talks about addresses as seen from the CPU, and I can see
> > two logical interpretations how the RAM is supposed to be visible from a device:
> > either all RAM would be visible contiguously at DMA address zero, or everything
> > would be visible at the same physical address as the CPU sees it.
> > 
> > If anyone picks the first interpretation, we will have to implement that
> > in Linux. We can of course hope that all hardware designs follow the second
> > interpretation, which would be more convenient for us here.
> 
> Indeed. Hope though we might, I would not be surprised to see a platform
> that does the first. In that case we could probably handle it with a
> ranges property that is DMA-controller facing instead of device facing.
> That would be able to handle the translation between CPU addressing and
> DMA addressing.
> 
> Come to think of it, doesn't PCI DMA have to deal with that situation if
> the PCI window is not 1:1 mapped into the CPU address space?

I think all PCI buses we support so far only need a single entry in the
dma-ranges property.

	Arnd