Re: [PATCH v22 22/23] KVM: arm64: CCA: Expose SVE VL register before VCPU finalization

[Suzuki K Poulose]

On 06/10/2026 06:35, Gavin Shan wrote:
> On 10/5/26 7:07 PM, Suzuki K Poulose wrote:
> > From: Jean-Philippe Brucker <jean-philippe@xxxxxxxxxx>
> >
> > Userspace must configure the SVE vector length before the Realm is 
> > created
> > (as it is part of the parameter for Realm creation), but the Realm VCPUs
> > cannot be finalized until after the Realm Descriptor has been created.
> >
> > KVM_GET_REG_LIST currently rejects the unfinalized VCPUs, which prevents
> > the userspace from discovering and configuring the VLs for  the Realm.
> >
> > Allow KVM_GET_REG_LIST for unfinalized RECs and make the SVE register
> > enumeration handle the unfinalized case explicitly. i.e., only expose
> > KVM_REG_ARM64_SVE_VLS before SVE is finalized.
> >
> > One adverse side effect of this change is that a KVM_GET_REG_LIST call 
> > that
> > only probes for the array size will now succeed even if SVE is not
> > finalized, but that seems harmless since the following KVM_GET_REG_LIST
> > with the full array will fail.
> >
> > Signed-off-by: Jean-Philippe Brucker <jean-philippe@xxxxxxxxxx>
> > Signed-off-by: Steven Price <steven.price@xxxxxxx>
> > Signed-off-by: Suzuki K Poulose <suzuki.poulose@xxxxxxx>
> > ---
> >   arch/arm64/kvm/arm.c   | 15 ++++++++++++++-
> >   arch/arm64/kvm/guest.c | 10 +++++-----
> >   2 files changed, 19 insertions(+), 6 deletions(-)
> >
> > diff --git a/arch/arm64/kvm/arm.c b/arch/arm64/kvm/arm.c
> > index fe707a0c47308..d99e7818f5894 100644
> > --- a/arch/arm64/kvm/arm.c
> > +++ b/arch/arm64/kvm/arm.c
> > @@ -2004,6 +2004,19 @@ static int kvm_arm_vcpu_set_events(struct 
> > kvm_vcpu *vcpu,
> >       return __kvm_arm_vcpu_set_events(vcpu, events);
> >   }
> > +/*
> > + * Realm VCPUs can be finalized only after the Realm descriptor is 
> > created.
> > + * But in order to seal the SVE VL, we need to allow the userspace to 
> > read/write
> > + * to the SVE_VL, before everything is finalized.
> > + * Allow the register list for RECs before the VCPUs are finalized.
> > + */
> > +static bool kvm_arm_vcpu_reg_list_allowed(struct kvm_vcpu *vcpu)
> > +{
> > +    if (kvm_arm_vcpu_is_finalized(vcpu))
> > +        return true;
> > +    return vcpu_is_rec(vcpu);
> > +}
> > +
>
> Needn't to keep this helper, and the code can be integrated to 
> kvm_arch_vcpu_ioctl(),
> see below.
>
> >   long kvm_arch_vcpu_ioctl(struct file *filp,
> >                unsigned int ioctl, unsigned long arg)
> >   {
> > @@ -2059,7 +2072,7 @@ long kvm_arch_vcpu_ioctl(struct file *filp,
> >               break;
> >           r = -EPERM;
> > -        if (!kvm_arm_vcpu_is_finalized(vcpu))
> > +        if (!kvm_arm_vcpu_reg_list_allowed(vcpu))
> >               break;
>
> Needn't to keep the helper kvm_arm_vcpu_reg_list_allowed() after its 
> logic is
> combined to kvm_arch_vcpu_ioctl().
>
>          /*
>           * Realm vCPUs can be finalized only after the Realm descriptor 
> is created.
>           * We need to allow access KVM_REG_ARM64_SVE_VLS before that so 
> that the
>           * register can be sealed.
>           */
>          if (!(vcpu_is_rec(vcpu) || kvm_arm_vcpu_is_finalized(vcpu))
>              break;


Wanted to keep the ioctl handling section cleaner and easier to read.
Hence the wrapper. The name is intuitive enough and compiler can do
away with inlining.

> >           r = -EFAULT;
> > diff --git a/arch/arm64/kvm/guest.c b/arch/arm64/kvm/guest.c
> > index b01d6622b8720..c3ca369882273 100644
> > --- a/arch/arm64/kvm/guest.c
> > +++ b/arch/arm64/kvm/guest.c
> > @@ -598,8 +598,8 @@ static unsigned long num_sve_regs(const struct 
> > kvm_vcpu *vcpu)
> >       if (!vcpu_has_sve(vcpu))
> >           return 0;
> > -    /* Policed by KVM_GET_REG_LIST: */
> > -    WARN_ON(!kvm_arm_vcpu_sve_finalized(vcpu));
> > +    if (!kvm_arm_vcpu_sve_finalized(vcpu))
> > +        return 1; /* KVM_REG_ARM64_SVE_VLS */
>
> Question: After a (Rec) vCPU is finalized, the returned number of 
> registers won't
> be 1. Is this expected? Otherwise, we need to use vcpu_is_rec() here.
>
>      /*
>       * KVM_REG_ARM64_SVE_VLS is visible for realm vCPUs no matter if
>       * they have been finalized.
>       */
>      if (vcpu_is_rec(vcpu))
>          return 1;

This is addressed in the next patch.

Cheers
Suzuki