Re: [PATCH 1/2] KVM: selftest: arm64: Support 5-level paging in stage 1 translation table

[Itaru Kitayama <itaru.kitayama@fujitsu.com>]

Hi Fuad,
On Fri, Sep 04, 2026 at 09:38:39AM +0100, Fuad Tabba wrote:
> Hi Itaru,
> 
> The usual subject prefix in the tree is "KVM: arm64: selftests:", for
> this patch and the next one.

Sure. Fix in v2.
> On Tue, 25 Aug 2026 at 22:18, Itaru Kitayama <itaru.kitayama@fujitsu.com> wrote:
> >
> > Add p4d_index() for when 5-level paging required, i.e., V52 guest mode
> > IDs. With the index helper function, _virt_pg_map() handles 5-level
> 
> Only VM_MODE_P52V52_4K has five levels. The 16K and 64K V52 modes get
> four and three, so this is for the 4K granule rather than for V52
> modes as a class.

Yes, understood. Will update the commit log.

> 
> ...
> 
> >  u64 *virt_get_pte_hva_at_level(struct kvm_vm *vm, gva_t gva, int level)
> >  {
> > +       int start_level = 4 - vm->mmu.pgtable_levels;
> >         u64 *ptep;
> >
> > +       TEST_ASSERT(level >= start_level && level <= 3,
> > +                   "Invalid translation level %d, valid range is %d-3",
> > +                   level, start_level);
> > +
> >         if (!vm->mmu.pgd_created)
> >                 goto unmapped_gva;
> >
> >         ptep = addr_gpa2hva(vm, vm->mmu.pgd) + pgd_index(vm, gva) * 8;
> >         if (!ptep)
> >                 goto unmapped_gva;
> > -       if (level == 0)
> > +       /*
> > +        * Stage-1 translation starts at level -1 for a five-level page
> > +        * table, and at levels 0, 1, or 2 for four-, three-, or two-level
> > +        * page tables, respectively.
> > +        */
> > +       if (level == start_level)
> 
> 
> This changes what level means for the existing three-level modes:
> level 0 used to return the top-level entry and now trips the assert,
> level 1 used to return the leaf and now returns the top level. No
> caller in tree asks for either of those two levels, and the new
> numbering matches the architecture.

Yes, as you checked this change won't break current tests, so if you
have better comments I will replace it with yours.

> 
> ...
> 
> > @@ -321,12 +356,14 @@ void aarch64_vcpu_setup(struct kvm_vcpu *vcpu, struct kvm_vcpu_init *init)
> >         case VM_MODE_PXXVYY_4K:
> >                 TEST_FAIL("AArch64 does not support 4K sized pages "
> >                           "with ANY-bit physical address ranges");
> > +       case VM_MODE_P52V52_64K:
> 
> These VM_MODE_ enumerators arrive in patch 2, so this patch does not
> build on its own:
> 
>   lib/arm64/processor.c:359:7: error: use of undeclared identifier
> 'VM_MODE_P52V52_64K'; did you mean 'VM_MODE_P52V48_64K'?
>   lib/arm64/processor.c:360:7: error: duplicate case value 'VM_MODE_P52V48_64K'
> 
> There are other build errors from the same cause. Could you move the
> enum values and their vm_guest_mode_string()/vm_guest_mode_params[]
> entries into this patch, or a new one altogether?

Yes, I have rearranged the series to avoid this.

> 
> >         case VM_MODE_P52V48_64K:
> >         case VM_MODE_P48V48_64K:
> >         case VM_MODE_P40V48_64K:
> >         case VM_MODE_P36V48_64K:
> >                 tcr_el1 |= TCR_TG0_64K;
> >                 break;
> > +       case VM_MODE_P52V52_16K:
> >         case VM_MODE_P52V48_16K:
> >         case VM_MODE_P48V48_16K:
> >         case VM_MODE_P40V48_16K:
> > @@ -334,6 +371,7 @@ void aarch64_vcpu_setup(struct kvm_vcpu *vcpu, struct kvm_vcpu_init *init)
> >         case VM_MODE_P36V47_16K:
> >                 tcr_el1 |= TCR_TG0_16K;
> >                 break;
> > +       case VM_MODE_P52V52_4K:
> >         case VM_MODE_P52V48_4K:
> >         case VM_MODE_P48V48_4K:
> >         case VM_MODE_P40V48_4K:
> > @@ -348,6 +386,9 @@ void aarch64_vcpu_setup(struct kvm_vcpu *vcpu, struct kvm_vcpu_init *init)
> >
> >         /* Configure output size */
> >         switch (vm->mode) {
> > +       case VM_MODE_P52V52_4K:
> > +       case VM_MODE_P52V52_16K:
> > +       case VM_MODE_P52V52_64K:
> >         case VM_MODE_P52V48_4K:
> >         case VM_MODE_P52V48_16K:
> >         case VM_MODE_P52V48_64K:
> > @@ -578,6 +619,42 @@ static u32 max_ipa_for_page_size(u32 vm_ipa, u32 gran,
> >                 return min(vm_ipa, 48U);
> >  }
> >
> > +u32 aarch64_get_supported_va_size(void)
> 
> This repeats aarch64_get_supported_page_sizes() below it and builds a
> second probe VM one line after the first. Could it take a u32 *va
> out-param, so both ID registers come off the one vCPU?
> 
> VARange is the 52-bit VA indicator for the 64K granule only, so a bare
> 52 or 48 reads as granule-independent. The caller in patch 2 only uses
> it for 64K, so nothing is wrong today.

Yes, I've dropped the _va_size() function and instead, expanded a bit
as you suggested the aarch64_get_supported_page_sizes() to check if vcpu
can address upto 52-bit VA space or not (48-bit max). 

Thanks,
Itaru.
> 
> Cheers,
> /fuad
> 
> 
> > +{
> > +       struct kvm_vcpu_init preferred_init = {};
> > +       int kvm_fd, vm_fd, vcpu_fd, err;
> > +       u64 val;
> > +       u32 va_range;
> > +       struct kvm_one_reg reg = {
> > +               .id     = KVM_ARM64_SYS_REG(SYS_ID_AA64MMFR2_EL1),
> > +               .addr   = (u64)&val,
> > +       };
> > +
> > +       kvm_fd = open_kvm_dev_path_or_exit();
> > +       vm_fd = __kvm_ioctl(kvm_fd, KVM_CREATE_VM, NULL);
> > +       TEST_ASSERT(vm_fd >= 0, KVM_IOCTL_ERROR(KVM_CREATE_VM, vm_fd));
> > +
> > +       vcpu_fd = ioctl(vm_fd, KVM_CREATE_VCPU, 0);
> > +       TEST_ASSERT(vcpu_fd >= 0, KVM_IOCTL_ERROR(KVM_CREATE_VCPU, vcpu_fd));
> > +
> > +       err = ioctl(vm_fd, KVM_ARM_PREFERRED_TARGET, &preferred_init);
> > +       TEST_ASSERT(err == 0, KVM_IOCTL_ERROR(KVM_ARM_PREFERRED_TARGET, err));
> > +
> > +       err = ioctl(vcpu_fd, KVM_ARM_VCPU_INIT, &preferred_init);
> > +       TEST_ASSERT(err == 0, KVM_IOCTL_ERROR(KVM_ARM_VCPU_INIT, err));
> > +
> > +       err = ioctl(vcpu_fd, KVM_GET_ONE_REG, &reg);
> > +       TEST_ASSERT(err == 0, KVM_IOCTL_ERROR(KVM_GET_ONE_REG, err));
> > +
> > +       va_range = FIELD_GET(ID_AA64MMFR2_EL1_VARange, val);
> > +
> > +       close(vcpu_fd);
> > +       close(vm_fd);
> > +       close(kvm_fd);
> > +
> > +       return va_range >= ID_AA64MMFR2_EL1_VARange_52 ? 52 : 48;
> > +}
> > +
> >  void aarch64_get_supported_page_sizes(u32 ipa, u32 *ipa4k,
> >                                       u32 *ipa16k, u32 *ipa64k)
> >  {
> >
> > --
> > 2.43.0
> >