Re: [PATCH v4 11/17] KVM: arm64: Filter out non-kernel addresses in kern_hyp_va

[Vincent Donnefort <vdonnefort@google.com>]

On Tue, Aug 18, 2026 at 03:45:23PM +0100, Fuad Tabba wrote:
> Hi Vincent,
> 
> 
> On Fri, 31 Jul 2026 at 15:36, Vincent Donnefort <vdonnefort@google.com> wrote:
> >
> > kern_hyp_va() is idempotent for the hypervisor linear space. This is
> > handy for nVHE hypervisor callers handling kvm_vcpu or kvm_arch
> > pointers. Those pointers can originate from the hypervisor space (when
> > protected mode is enabled, we don't trust the kernel and the hypervisor
> > uses its own copy) or from the kernel space (we do trust the kernel in
> > "non-protected" nVHE).
> >
> > This idempotence does not hold for addresses within the hypervisor
> > private range, like the ones you get from the pKVM heap allocator
> > (hyp_alloc()). To resolve this, filter out non-kernel addresses based on
> > PAGE_OFFSET.
> >
> > Leave the assembly version untouched as it has no current users.
> >
> > Reviewed-by: Fuad Tabba <fuad.tabba@linux.dev>
> > Tested-by: Fuad Tabba <fuad.tabba@linux.dev>
> > Signed-off-by: Vincent Donnefort <vdonnefort@google.com>
> >
> > diff --git a/arch/arm64/include/asm/kvm_mmu.h b/arch/arm64/include/asm/kvm_mmu.h
> > index 6eae7e7e2a68..d60e5f2de10c 100644
> > --- a/arch/arm64/include/asm/kvm_mmu.h
> > +++ b/arch/arm64/include/asm/kvm_mmu.h
> > @@ -126,6 +126,9 @@ static __always_inline unsigned long __kern_hyp_va(unsigned long v)
> >   * replace the instructions with `nop`s.
> >   */
> >  #ifndef __KVM_VHE_HYPERVISOR__
> > +       if (!is_ttbr1_addr(v))
> > +               return v;
> > +
> 
> I gave this a tag earlier and only came back to the cost side now, sorry.
> 
> The check is unconditional, so plain nVHE pays it on every
> kern_hyp_va(), including the one in __kvm_vcpu_run(), and it has no
> private range to protect.
> 
> Could it be gated on is_protected_kvm_enabled(), the same way patch 17
> gates hyp_trace_buffer_alloc_bpages()?
> 
> Cheers,
> /fuad

Sure, that's really a micro optimisation though, I don't think that
is_ttbr1_addr() is expensive.

-- 
Vincent
> 
> 
> >         asm volatile(ALTERNATIVE_CB("and %0, %0, #1\n"         /* mask with va_mask */
> >                                     "ror %0, %0, #1\n"         /* rotate to the first tag bit */
> >                                     "add %0, %0, #0\n"         /* insert the low 12 bits of the tag */
> > --
> > 2.55.0.508.g3f0d502094-goog
> >