On Sat, Jul 09, 2011 at 09:33:06AM -0500, Rob Herring wrote:

I noticed this is going to have conflicts with your ARM_DMA_ZONE_SIZE
series. Can you add this into your series? This still presents another
issue to solve for single kernel binary, but hopefully the dma mapping
work will help.

That's a different problem - its about limiting the maximum DMA mask
which can be set, so that memory is sourced from the right place
even for drivers which use 32-bit masks.

It's needed to ensure that dma_set_coherent_mask() can be overridden.
One solution to it would be to ensure that we always define
ARCH_HAS_DMA_SET_COHERENT_MASK, and handle the differences internally,
which is something I've been thinking about adding to my DMA bounce
series of patches.