Re: [RFC PATCH v1 2/8] KVM: selftests: x86: Support guest running on canonical linear-address organization

[Sean Christopherson]

On Thu, Nov 02, 2023, Zeng Guang wrote:
> Setup execution environment running on 64-bit linear addresses for
> user and supervisor mode.
> 
> Define the linear address based on 48-bit canonical format in which
> bits 63:47 of the address are identical. All addresses to system data
> structure are shifted to supervisor-mode address space.
> 
> Extend page table mapping for supervisor mode to same guest physical
> address. This allows guest in supervisor mode can run in the
> corresponding canonical linear address space.
> 
> Signed-off-by: Zeng Guang <guang.zeng@xxxxxxxxx>
> ---
>  .../selftests/kvm/include/x86_64/processor.h  |  6 ++++
>  tools/testing/selftests/kvm/lib/kvm_util.c    |  6 ++--
>  .../selftests/kvm/lib/x86_64/processor.c      | 28 ++++++++++++-------
>  3 files changed, 28 insertions(+), 12 deletions(-)
> 
> diff --git a/tools/testing/selftests/kvm/include/x86_64/processor.h b/tools/testing/selftests/kvm/include/x86_64/processor.h
> index 25bc61dac5fb..00f7337a520a 100644
> --- a/tools/testing/selftests/kvm/include/x86_64/processor.h
> +++ b/tools/testing/selftests/kvm/include/x86_64/processor.h
> @@ -1256,4 +1256,10 @@ void virt_map_level(struct kvm_vm *vm, uint64_t vaddr, uint64_t paddr,
>  #define PFERR_GUEST_PAGE_MASK	BIT_ULL(PFERR_GUEST_PAGE_BIT)
>  #define PFERR_IMPLICIT_ACCESS	BIT_ULL(PFERR_IMPLICIT_ACCESS_BIT)
>  
> +/*
> + * X86 kernel linear address defines
> + */
> +#define KERNEL_LNA_OFFSET 0xffff800000000000

Please don't make up acronyms, I can more or less glean what LNA is from the
context _here_, but in other usage I would truly have no idea.

> diff --git a/tools/testing/selftests/kvm/lib/x86_64/processor.c b/tools/testing/selftests/kvm/lib/x86_64/processor.c
> index 9f4b8c47edce..6f4295a13d00 100644
> --- a/tools/testing/selftests/kvm/lib/x86_64/processor.c
> +++ b/tools/testing/selftests/kvm/lib/x86_64/processor.c
> @@ -227,6 +227,13 @@ void __virt_pg_map(struct kvm_vm *vm, uint64_t vaddr, uint64_t paddr, int level)
>  void virt_arch_pg_map(struct kvm_vm *vm, uint64_t vaddr, uint64_t paddr)
>  {
>  	__virt_pg_map(vm, vaddr, paddr, PG_LEVEL_4K);
> +
> +	/*
> +	 * Map same paddr to kernel linear address space. Make execution
> +	 * environment supporting running both in user and kernel mode.
> +	 */
> +	if (!(vaddr & BIT_ULL(63)))
> +		__virt_pg_map(vm, (uint64_t)KERNEL_ADDR(vaddr), paddr, PG_LEVEL_4K);

I really don't like the idea of piling hacks on top of selftests' misguided
infrastructure.  Letting tests control virtual addresses is all kinds of stupid.
Except for ARM's ucall_arch_init(), I don't think there's a single user of
virt_map() that _needs_ a specific address, e.g. most tests just identity map
the GPA.

So rather than fudge things by stuffing two mappings, which is wasteful for 99%
of mappings and will likely be a maintenance nightmare, I think we should go
straight to getting x86's kernel mappings setup correctly from time zero.

>From KUT experience, using USER mappings for kernel accesses is explosions waiting
to happen due to SMAP and SMEP.  And expecting developers to remember to sprinkle
KERNEL_ADDR() everywhere is not remotely maintainable.

In other words, give virt_arch_pg_map() (or ideally, the common virt_map()) over
picking the virtual address, and then plumb in information as to whether the
allocation is USER vs. SUPERVISOR.