Re: [PATCH] fb.h: ARM uses __raw_{read/write}

On Mon, Jun 13, 2011 at 03:37:10PM +0200, Arnd Bergmann wrote:

On Monday 13 June 2011 12:08:19 Catalin Marinas wrote:

quoted

The change proposed by Hartley wouldn't make much difference from the
current volatile accesses (__raw_* accessors are implemented as volatile
on ARM).

I guess it would mainly make a difference if we get a platform where
the PCI bus window is not in the same address space as the regular
physical memory and the accessors actually need to do a computation
on the address. AFAICT, ARM does not currently have any such platform
and I would hope that it says that way.

Actually we do.

Footbridge has this situation: PCI memory space is 0 to 4GB, and appears
in physical space at 2GB-4GB physical, and can be switched between the
low and high 2GB of PCI space.

Non-PCI peripherals are located at 0 to 2GB physical.

We handle this by having pcibios_bus_to_resource() and
pcibios_resource_to_bus(), which translates from the bus space to a
cookie suitable for ioremap() and /proc/iomem.  This cookie happens to
be the physical address.

We avoid the obvious problem with the upper 2GB of PCI memory space by
placing the system RAM at 0xe0000000 on the PCI bus, and not allocating
any peripherals in the upper 2GB of PCI memory space.

So, I rather wish that people would stop saying that "ioremap takes the
PCI bus address".  It doesn't - it takes a platform specific cookie to
allow a mapping to be setup, which in the case of pci will be a cookie
provided from the PCI bus address by the pcibios_bus_to_resource()
platform code.

And the final thing to note on this is that fbmem/fbcon does work - see
the Cyber2000/CyberPro VGA driver which has been - and still is - used
on Netwinder (footbridge).