On Wed, 2009-08-26 at 15:20 -0500, Becky Bruce wrote:

On Aug 26, 2009, at 9:08 AM, Michael Ellerman wrote:

quoted

On Wed, 2009-08-26 at 22:29 +1000, Benjamin Herrenschmidt wrote:

quoted

On Mon, 2009-08-24 at 21:48 +0200, Christoph Hellwig wrote:

quoted

On Mon, Aug 24, 2009 at 11:17:14AM -0500, Becky Bruce wrote:

quoted

Previously, this was specified as a void *, but that's not
large enough on 32-bit systems with 36-bit physical
addressing support.  Change the type to dma_addr_t so it
will scale based on the size of a dma address.

This looks extreml ugly to me.  It seems like the typical use is to
store a pointer to a structure.  So what about making the direct
dma case follow that general scheme instead?

E.g. declare a

struct direct_dma_data {
	dma_addr_t	direct_dma_offset;
};

and have one normal instace of it, and one per weird cell device.

Right, but we want to avoid a structure for the classic case of 32- 
bit
systems with no iommu...

I wouldn't mind doing a union here.

That might be best, the patch as it stands is a horrible mess of  
casts.

Let's be fair - the code before was a horrible mess of casts, I've  
just moved them :)

Yeah true. Though I think we end up with more casts because there were
more call sites using it as a pointer originally. But yeah it's not
pretty either way.

quoted

Stashing a dma_addr_t into a void * is sort of gross, but storing a
pointer to some struct (a void *) in a dma_addr_t is _really_ gross :)

Both are revolting (and storing a dma_addr_t into a void * is really  
gross when the void * is smaller than the dma_addr_t!!).  A union  
might not be a bad idea, though.  I'll look at doing that instead.

Cool. That is how we're using it, sometimes it points to something
sometimes it's a dma_addr_t, so I think a union will work.

cheers