On 04/05/2018 11:30, Marc Zyngier wrote:
>> you should also add the arm/mips/power maintainers so that they can review their
>> struct kvm_arch if any of the content is passed directly to the hardware (e.g. with
>> __pa). A quick looks seems to indicate that there is no issue.
> Well, there is a small issue, at least on arm64. Only VAs that are part
> of the linear mapping can easily be mapped at EL2 (so that we can play
> our kern_to_hyp_va trick on ARMv8.0 cores), and the kvm structure is
> assumed to be physically contiguous.
>
> Switch to a vmalloc'ed structure breaks these two requirements since, as
> you noticed it, the kvm struct is more than a single page.
>
> I'd suggest you let each architecture buy into this, or provide a
> kmalloc-based kvm_arch_{alloc,free}_vm for architectures that cannot use
> the vmalloc area for this. On arm64, we could use vmalloc on VHE-capable
> cores, and stick to kmalloc for the rest.
>
Alternatively you could have a #define __KVM_KMALLOC_STRUCT_KVM and
define it in arch/arm*/include/asm/kvm_host.h.
Paolo
