Re: [PATCH v4 3/6] KVM: arm64: Use per guest ID register for ID_AA64PFR0_EL1.[CSV2|CSV3]

[Fuad Tabba <tabba@google.com>]

Hi,

On Fri, Mar 17, 2023 at 5:06 AM Jing Zhang <jingzhangos@google.com> wrote:
>
> With per guest ID registers, ID_AA64PFR0_EL1.[CSV2|CSV3] settings from
> userspace can be stored in its corresponding ID register.
>
> No functional change intended.
>
> Signed-off-by: Jing Zhang <jingzhangos@google.com>
> ---
>  arch/arm64/include/asm/kvm_host.h  |  2 --
>  arch/arm64/kvm/arm.c               | 19 +------------------
>  arch/arm64/kvm/hyp/nvhe/sys_regs.c |  7 +++----
>  arch/arm64/kvm/id_regs.c           | 30 ++++++++++++++++++++++--------
>  4 files changed, 26 insertions(+), 32 deletions(-)
>
> diff --git a/arch/arm64/include/asm/kvm_host.h b/arch/arm64/include/asm/kvm_host.h
> index fb6b50b1f111..e926ea91a73c 100644
> --- a/arch/arm64/include/asm/kvm_host.h
> +++ b/arch/arm64/include/asm/kvm_host.h
> @@ -230,8 +230,6 @@ struct kvm_arch {
>
>         cpumask_var_t supported_cpus;
>
> -       u8 pfr0_csv2;
> -       u8 pfr0_csv3;
>         struct {
>                 u8 imp:4;
>                 u8 unimp:4;
> diff --git a/arch/arm64/kvm/arm.c b/arch/arm64/kvm/arm.c
> index 4579c878ab30..c78d68d011cb 100644
> --- a/arch/arm64/kvm/arm.c
> +++ b/arch/arm64/kvm/arm.c
> @@ -104,22 +104,6 @@ static int kvm_arm_default_max_vcpus(void)
>         return vgic_present ? kvm_vgic_get_max_vcpus() : KVM_MAX_VCPUS;
>  }
>
> -static void set_default_spectre(struct kvm *kvm)
> -{
> -       /*
> -        * The default is to expose CSV2 == 1 if the HW isn't affected.
> -        * Although this is a per-CPU feature, we make it global because
> -        * asymmetric systems are just a nuisance.
> -        *
> -        * Userspace can override this as long as it doesn't promise
> -        * the impossible.
> -        */
> -       if (arm64_get_spectre_v2_state() == SPECTRE_UNAFFECTED)
> -               kvm->arch.pfr0_csv2 = 1;
> -       if (arm64_get_meltdown_state() == SPECTRE_UNAFFECTED)
> -               kvm->arch.pfr0_csv3 = 1;
> -}
> -
>  /**
>   * kvm_arch_init_vm - initializes a VM data structure
>   * @kvm:       pointer to the KVM struct
> @@ -151,9 +135,8 @@ int kvm_arch_init_vm(struct kvm *kvm, unsigned long type)
>         /* The maximum number of VCPUs is limited by the host's GIC model */
>         kvm->max_vcpus = kvm_arm_default_max_vcpus();
>
> -       set_default_spectre(kvm);
> -       kvm_arm_init_hypercalls(kvm);
>         kvm_arm_set_default_id_regs(kvm);
> +       kvm_arm_init_hypercalls(kvm);
>
>         /*
>          * Initialise the default PMUver before there is a chance to
> diff --git a/arch/arm64/kvm/hyp/nvhe/sys_regs.c b/arch/arm64/kvm/hyp/nvhe/sys_regs.c
> index 08d2b004f4b7..0e1988740a65 100644
> --- a/arch/arm64/kvm/hyp/nvhe/sys_regs.c
> +++ b/arch/arm64/kvm/hyp/nvhe/sys_regs.c
> @@ -93,10 +93,9 @@ static u64 get_pvm_id_aa64pfr0(const struct kvm_vcpu *vcpu)
>                 PVM_ID_AA64PFR0_RESTRICT_UNSIGNED);
>
>         /* Spectre and Meltdown mitigation in KVM */
> -       set_mask |= FIELD_PREP(ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV2),
> -                              (u64)kvm->arch.pfr0_csv2);
> -       set_mask |= FIELD_PREP(ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV3),
> -                              (u64)kvm->arch.pfr0_csv3);
> +       set_mask |= vcpu->kvm->arch.id_regs[IDREG_IDX(SYS_ID_AA64PFR0_EL1)] &
> +               (ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV2) |
> +                       ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV3));

This triggers a compiler warning now since the variable `struct kvm
*kvm` isn't used anymore, this, however, isn't the main issue.

The main issue is that `struct kvm` here (vcpu->kvm) is the
hypervisor's version for protected vms, and not the host's. Therefore,
reading that value is wrong. That said, this is an existing bug in
pKVM since kvm->arch.pfr0_csv2 and kvm->arch.pfr0_csv3 are not
initialized.

The solution would be to track the spectre/meltown state at hyp and
use that. I'll submit a patch that does that. In the meantime, I think
that it would be better not to set the CSV bits for protected VMs,
which is the current behavior in practice.

Non-protected VMs in protected mode go back to the host on id register
traps, and use the host's `struct kvm`, so they should behave as
expected.

Thanks,
/fuad


>
>         return (id_aa64pfr0_el1_sys_val & allow_mask) | set_mask;
>  }
> diff --git a/arch/arm64/kvm/id_regs.c b/arch/arm64/kvm/id_regs.c
> index e393b5730557..b60ca1058301 100644
> --- a/arch/arm64/kvm/id_regs.c
> +++ b/arch/arm64/kvm/id_regs.c
> @@ -61,12 +61,6 @@ u64 kvm_arm_read_id_reg(const struct kvm_vcpu *vcpu, u32 id)
>                 if (!vcpu_has_sve(vcpu))
>                         val &= ~ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_SVE);
>                 val &= ~ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_AMU);
> -               val &= ~ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV2);
> -               val |= FIELD_PREP(ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV2),
> -                                 (u64)vcpu->kvm->arch.pfr0_csv2);
> -               val &= ~ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV3);
> -               val |= FIELD_PREP(ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV3),
> -                                 (u64)vcpu->kvm->arch.pfr0_csv3);
>                 if (kvm_vgic_global_state.type == VGIC_V3) {
>                         val &= ~ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_GIC);
>                         val |= FIELD_PREP(ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_GIC), 1);
> @@ -201,6 +195,7 @@ static int set_id_aa64pfr0_el1(struct kvm_vcpu *vcpu,
>                                u64 val)
>  {
>         u8 csv2, csv3;
> +       u64 sval = val;
>
>         /*
>          * Allow AA64PFR0_EL1.CSV2 to be set from userspace as long as
> @@ -225,8 +220,7 @@ static int set_id_aa64pfr0_el1(struct kvm_vcpu *vcpu,
>         if (val)
>                 return -EINVAL;
>
> -       vcpu->kvm->arch.pfr0_csv2 = csv2;
> -       vcpu->kvm->arch.pfr0_csv3 = csv3;
> +       vcpu->kvm->arch.id_regs[IDREG_IDX(reg_to_encoding(rd))] = sval;
>
>         return 0;
>  }
> @@ -529,4 +523,24 @@ void kvm_arm_set_default_id_regs(struct kvm *kvm)
>                 val = read_sanitised_ftr_reg(id);
>                 kvm->arch.id_regs[IDREG_IDX(id)] = val;
>         }
> +       /*
> +        * The default is to expose CSV2 == 1 if the HW isn't affected.
> +        * Although this is a per-CPU feature, we make it global because
> +        * asymmetric systems are just a nuisance.
> +        *
> +        * Userspace can override this as long as it doesn't promise
> +        * the impossible.
> +        */
> +       val = kvm->arch.id_regs[IDREG_IDX(SYS_ID_AA64PFR0_EL1)];
> +
> +       if (arm64_get_spectre_v2_state() == SPECTRE_UNAFFECTED) {
> +               val &= ~ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV2);
> +               val |= FIELD_PREP(ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV2), 1);
> +       }
> +       if (arm64_get_meltdown_state() == SPECTRE_UNAFFECTED) {
> +               val &= ~ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV3);
> +               val |= FIELD_PREP(ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV3), 1);
> +       }
> +
> +       kvm->arch.id_regs[IDREG_IDX(SYS_ID_AA64PFR0_EL1)] = val;
>  }
> --
> 2.40.0.rc1.284.g88254d51c5-goog
>