Re: [PATCH v1 04/10] KVM: arm64: nv: Return a failed stage-2 descriptor read as a fault

[Oliver Upton <oupton@kernel.org>]

On Mon, Sep 28, 2026 at 04:25:01PM +0100, Fuad Tabba wrote:
> When KVM walks an L1 guest's stage-2 tables and can't read a
> descriptor, for example because the guest points a table at an IPA with
> no memslot, it sets a synchronous external abort on the walk but
> returns the -EFAULT from kvm_read_guest(). As with an invalid
> VTCR_EL2, KVM then leaves AT S12E{0,1}{R,W} unretired, so the guest
> retries it forever. On an L2 abort, KVM injects the abort and
> returns -EFAULT from KVM_RUN.
> 
> Return 1, as the walk already does when it can't update a descriptor,
> leaving -EAGAIN from a descriptor race as its only negative return.
> The AT then retires with the external abort in PAR_EL1, which is what a
> stage-1 walk records for a descriptor it can't read.
> 
> Fixes: fd276e71d1e7b ("KVM: arm64: nv: Handle shadow stage 2 page faults")
> Fixes: 92c6443222ca4 ("KVM: arm64: Propagate PTW errors up to AT emulation")
> Signed-off-by: Fuad Tabba <fuad.tabba@linux.dev>
> ---
>  arch/arm64/kvm/nested.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/arch/arm64/kvm/nested.c b/arch/arm64/kvm/nested.c
> index a67be19e73250..f9a2e50c0b567 100644
> --- a/arch/arm64/kvm/nested.c
> +++ b/arch/arm64/kvm/nested.c
> @@ -304,7 +304,7 @@ static int walk_nested_s2_pgd(struct kvm_vcpu *vcpu, phys_addr_t ipa,
>  		ret = read_guest_s2_desc(vcpu, paddr, &desc, wi);
>  		if (ret < 0) {
>  			out->esr = ESR_ELx_FSC_SEA_TTW(level);
> -			return ret;
> +			return 1;

Hmm. Rather than fixing these things individually, I'd prefer aligning return
codes between the S1 and S2 walks. I agree with the approach; negative
return codes should be reserved for host-visible failure reasons whereas
everything guest visible gets communicated through s1_walk_result /
kvm_s2_trans.

Thanks,
Oliver