[PATCH 2.6.39 v2] V4L: videobuf-dma-contig: fix mmap_mapper broken on ARM

On Tue, 12 Apr 2011, Russell King - ARM Linux wrote:

On Tue, Apr 12, 2011 at 11:06:34PM +0200, Janusz Krzysztofik wrote:

quoted

The patch tries to solve this regression by using 
virt_to_phys(bus_to_virt(mem->dma_handle)) instead of problematic 
virt_to_phys(mem->vaddr).

Who says that DMA handles are bus addresses in the strictest sense?

DMA handles on ARM are the bus address to program 'dev' with in order
for it to access the memory mapped by dma_alloc_coherent().  On some
ARM platforms, this bus address is dependent on 'dev' - such as platforms
with more than one root PCI bus, and so bus_to_virt() just doesn't
hack it.

What is really needed is for this problem - the mapping of DMA coherent
memory into userspace - to be solved with a proper arch API rather than
all these horrible hacks which subsystems instead invent.  That's
something I tried to do with the dma_mmap_coherent() stuff but it was
shot down by linux-arch as (iirc) PA-RISC objected to it.

Hence why ARM only implements it.

Maybe the video drivers should try to resurect the idea, maybe only
allowing this support for architectures which provide dma_mmap_coherent().

I do not know how this fits into the present discussion. Perhaps everyone 
who reads the above message is well aware of what is below. If so my 
comment below is superfluous. But just in case things are otherwise it 
might save someone a bit of trouble in trying to write something which 
will work "everywhere":

If one is speaking here of architecture problems, there is the additional 
problem that some ARM systems might have not two PCI buses, but instead 
no PCI bus at all.

Theodore Kilgore