[RFC PATCH 0/2] arm64: Add support for 48-bit Physical Addresses

[Catalin Marinas]

On Tue, Dec 10, 2013 at 04:53:24PM +0000, Radha Mohan wrote:
> On Thu, Dec 5, 2013 at 11:00 PM, Catalin Marinas
> <catalin.marinas at arm.com> wrote:
> > On Thu, Dec 05, 2013 at 02:35:52PM +0000, Radha Mohan wrote:
> >> Before going for a next version of patch to fix breaking SMMU and KVM,
> >> I would like to know  about the best possible implementation.
> >> Should I go for a compile time option? I am not in favor of this. What
> >> are options if in future there's 16KB page support in the kernel.
> >
> > As I said in the other email, separate the shifts definitions (primarily
> > page size dependent) from the type definitions (dependent on the number
> > of levels).
> >
> > Just to understand your requirements better. Does your platform have
> > more than 256GB of RAM? With the current configuration, the VA space is
> > 39-bit with 4K (which would allow 512GB minus some vmalloc and modules
> > space) and 42-bit with 64K pages (2TB minus vmalloc/modules).
> 
> Sorry for the delay in responding.
> More than 256GB DRAM is a likely possibility. But I am unsure of the
> exact page size that would be used. That is the reason for expanding
> the VA bits to full and have flexibility for customers. As Marc
> mentioned this is a rare requirement for now in ARMv8 processors, but
> would be nice to have a plan for long term. I would prefer a runtime
> support for this i.e reading the ID_AA64MMFR0_EL1.PARange value but I
> think this would be a major rework.

Don't confuse the physical address range (given by PARange) with the
virtual address range. On arm64 we don't have to use highmem to access
the RAM, so if you have lots of RAM (maximum architecture limit with
3-levels and 4K pages is 512GB) we need to expand the VA space to map
it. But just having 48-bit PA space doesn't necessarily require the same
VA space (unless for some weird reason all the RAM lives beyond bit 39
which would require the idmap to have 4 levels, but even then we could
do this just for a static idmap table).

-- 
Catalin