disabling MMU of ARMv7-A

[Dave Martin]

On Fri, Mar 28, 2014 at 10:43:49AM +0000, Catalin Marinas wrote:
> On Fri, Mar 28, 2014 at 07:15:28AM +0000, Chanho Park wrote:
> > I want to know is there anyone who tried disabling MMU of ARMv7-A likes
> > cortex-a9.
> > Yes. I know this is a silly question :)
> > I turned off CONFIG_MMU option of my exynos4 board and specified
> > CONFIG_DRAM_BASE/SIZE and REMAP_VECTORS_TO_RAM.
> > The problem is occurred in here there is no way to specify the vector table
> > address.
> > As I know the cortex-M3/4 has VTOR(Vector table offset register) to specify
> > the vector table address.
> > However, ARMv7-A/R doesn't have such register. They can only control whether
> > use _HIGH_ vector or not.
> > In this situation, can I find a way to boot successfully with
> > nommu+cortex-a9?
> 
> There is no way to set a random vectors address on A9. But even if you

Picking up an old thread here, but just to clarify:

I think the vectors can be moved on A9 using the VBAR.  All v6K and
later processors with the Security Extensions (including v7-A, v8-R)
have this.  However, Linux still needs to work on CPUs that don't have
this feature.  This may include a lot of platforms capable of running
nommu kernels.

> want a !MMU kernel, I would not recommend disabling the MMU entirely
> because your memory accesses are strongly ordered (and slow and
> unaligned accesses will not work). It's better to create some flat
> mapping, possibly before starting the kernel and keep the MMU enabled.
> Such mapping would contain a vectors page as well (low or high).

Agreed -- CONFIG_MMU=n does not need to imply that the MMU is turned off:
everything must be indentity-mapped (VA = PA), but it is still desirable
to define attributes.  For SMP, it's *necessary* to define attributes,
otherwise we can't guarantee working spinlocks etc.

I don't actually know whether nommu works with SMP on v7+-A today.  Maybe
someone else can comment on that.

Cheers
---Dave