[PATCH v7 07/12] KVM: arm64: PMU: Set PMCR_EL0.N for vCPU based on the associated PMU

[Oliver Upton]

On Tue, Oct 17, 2023 at 10:25:50AM -0700, Raghavendra Rao Ananta wrote:
> On Tue, Oct 17, 2023 at 10:09 AM Oliver Upton <oliver.upton at linux.dev> wrote:
> >
> > On Tue, Oct 17, 2023 at 09:58:08AM -0700, Raghavendra Rao Ananta wrote:
> > > On Mon, Oct 16, 2023 at 10:52 PM Oliver Upton <oliver.upton at linux.dev> wrote:
> > > >
> > > > On Mon, Oct 16, 2023 at 02:35:52PM -0700, Raghavendra Rao Ananta wrote:
> > > >
> > > > [...]
> > > >
> > > > > > What's the point of doing this in the first place? The implementation of
> > > > > > kvm_vcpu_read_pmcr() is populating PMCR_EL0.N using the VM-scoped value.
> > > > > >
> > > > > I guess originally the change replaced read_sysreg(pmcr_el0) with
> > > > > kvm_vcpu_read_pmcr(vcpu) to maintain consistency with others.
> > > > > But if you and Sebastian feel that it's an overkill and directly
> > > > > getting the value via vcpu->kvm->arch.pmcr_n is more readable, I'm
> > > > > happy to make the change.
> > > >
> > > > No, I'd rather you delete the line where PMCR_EL0.N altogether.
> > > > reset_pmcr() tries to initialize the field, but your
> > > > kvm_vcpu_read_pmcr() winds up replacing it with pmcr_n.
> > > >
> > > I didn't get this comment. We still do initialize pmcr, but using the
> > > pmcr.n read via kvm_vcpu_read_pmcr() instead of the actual system
> > > register.
> >
> > You have two bits of code trying to do the exact same thing:
> >
> >  1) reset_pmcr() initializes __vcpu_sys_reg(vcpu, PMCR_EL0) with the N
> >     field set up.
> >
> >  2) kvm_vcpu_read_pmcr() takes whatever is in __vcpu_sys_reg(vcpu, PMCR_EL0),
> >     *masks out* the N field and re-initializes it with vcpu->kvm->arch.pmcr_n
> >
> > Why do you need (1) if you do (2)?
> >
> Okay, I see what you mean now. In that case, let reset_pmcr():
> - Initialize 'pmcr' using  vcpu->kvm->arch.pmcr_n
> - Set ARMV8_PMU_PMCR_LC as appropriate in 'pmcr'
> - Write 'pmcr' to the vcpu reg
> 
> From here on out, kvm_vcpu_read_pmcr() would read off of this
> initialized value, unless of course, userspace updates the pmcr.n.
> Is this the flow that you were suggesting?

Just squash this in:

diff --git a/arch/arm64/kvm/sys_regs.c b/arch/arm64/kvm/sys_regs.c
index d1db1f292645..7b54c7843bef 100644
--- a/arch/arm64/kvm/sys_regs.c
+++ b/arch/arm64/kvm/sys_regs.c
@@ -743,10 +743,8 @@ static u64 reset_pmselr(struct kvm_vcpu *vcpu, const struct sys_reg_desc *r)
 
 static u64 reset_pmcr(struct kvm_vcpu *vcpu, const struct sys_reg_desc *r)
 {
-	u64 pmcr;
+	u64 pmcr = 0;
 
-	/* Only preserve PMCR_EL0.N, and reset the rest to 0 */
-	pmcr = kvm_vcpu_read_pmcr(vcpu) & (ARMV8_PMU_PMCR_N_MASK << ARMV8_PMU_PMCR_N_SHIFT);
 	if (!kvm_supports_32bit_el0())
 		pmcr |= ARMV8_PMU_PMCR_LC;
 

-- 
Thanks,
Oliver