[PATCH 13/26] ARM: pxa: use correct __iomem annotations

[Arnd Bergmann]

On Friday 07 October 2011, Eric Miao wrote:
> > @@ -18,11 +24,11 @@
> >  * peripherals on APB, let's count it into the ABP mapping area.
> >  */
> >  #define APB_PHYS_BASE          0xd4000000
> > -#define APB_VIRT_BASE          0xfe000000
> > +#define APB_VIRT_BASE          IOMEM(0xfe000000)
> 
> To be honest, I'd really like to keep the *_VIRT_BASE definitions to be
> type independent.
> 
> And have the actual register definitions to be casted to void __iomem *
> when being defined, e.g.
> 
> #define APBC_REG(x)     IOMEM(APBC_VIRT_BASE + (x))
> 
> #define APBC_UART1      APBC_REG(0x000)
> 
> Arnd, do we have some standard guidelines on this for all SoCs
> to follow? As I know, it's currently still being a mess.

We don't have any formal guidelines yet, but I'd really love to get
rid of all the arbitrary type casts to make use of the built-in
type checking of the compiler and sparse. A virtual base address
for registers is conventionally an __iomem pointer, so defining it
as something else is completely bogus. I have started making patches
for a number of platforms for this.

Ideally we should have very little code directly using, but the fact
is that they are there and we won't remove them all anytime soon, so
we should at least use the correct types here.

Another issue that goes together with this is that right now our
readl/writel macros accept any input type (pointer, __iomem pointer,
unsigned long, unsigned int), and I have a patch to make that stricter
but that requires fixing up all the places where we do a
readl(APBC_VIRT_BASE + x) that Russell mentioned.

The only place where this requires adding extra type casts right now is
the iotable setup, which I hope we can also fix eventually by splitting
the static I/O mapping setup from other static mappings (MT_MEMORY,
MT_MEMORY_NONCACHED, ...).

	Arnd