Re: Increase GICV size to 64KB so !4KB page-size kernels are able to initialize KVM

[Marc Zyngier <marc.zyngier@arm.com>]

On 22/06/16 07:48, Itaru Kitayama wrote:
> Against the tag v4.7-rc2 in the kvmtree.
> 
> ---
>   drivers/irqchip/irq-gic.c | 2 +-
>   1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/drivers/irqchip/irq-gic.c b/drivers/irqchip/irq-gic.c
> index fbc4ae2..bd463a3 100644
> --- a/drivers/irqchip/irq-gic.c
> +++ b/drivers/irqchip/irq-gic.c
> @@ -1425,7 +1425,7 @@ static bool __init gic_validate_dist(struct 
> acpi_subtable_header *header,
>   #define ACPI_GICV2_DIST_MEM_SIZE       (SZ_4K)
>   #define ACPI_GIC_CPU_IF_MEM_SIZE       (SZ_8K)
>   #define ACPI_GICV2_VCTRL_MEM_SIZE      (SZ_4K)
> -#define ACPI_GICV2_VCPU_MEM_SIZE       (SZ_8K)
> +#define ACPI_GICV2_VCPU_MEM_SIZE       (SZ_64K)
> 
>   static void __init gic_acpi_setup_kvm_info(void)
>   {
> 

I'm afraid this is the wrong approach. No matter how you look at it, the
size of the GICV region is 8kB, and not 64kB. On some systems the GICV
region is *aliased* over 128kB (first 4kB aliased over the first 64kB
page, second 4kB aliased over the second 64kB page). The usable range
for the GICV region is then the f000-10fff region in order to preserve
the contiguity of the two 4kB blocks.

In that case, ACPI ought to provide the base address of the usable
range, not some dummy address. It would be different if the ACPI tables
would provide the size of the region (where we could make an educated
guess at what is happening, just like with DT), but the information is
simply not available. So my advise here is to fix your ACPI tables (or
get them fixed).

What is actually missing in KVM is a way to describe the offset within a
page at which the GICV region must be mapped. I have patches for that
that I may revive if there is some interest.

Thanks,

	M.
-- 
Jazz is not dead. It just smells funny...