On Tue, Aug 22, 2023 at 11:26:23AM +0800, Shaoqin Huang wrote:
[...]
> > > > +static int set_pmcr(struct kvm_vcpu *vcpu, const struct sys_reg_desc *r,
> > > > + u64 val)
> > > > +{
> > > > + struct kvm *kvm = vcpu->kvm;
> > > > + u64 new_n, mutable_mask;
> > > > + int ret = 0;
> > > > +
> > > > + new_n = FIELD_GET(ARMV8_PMU_PMCR_N, val);
> > > > +
> > > > + mutex_lock(&kvm->arch.config_lock);
> > > > + if (unlikely(new_n != kvm->arch.pmcr_n)) {
> > > > + /*
> > > > + * The vCPU can't have more counters than the PMU
> > > > + * hardware implements.
> > > > + */
> > > > + if (new_n <= kvm->arch.pmcr_n_limit)
> > > > + kvm->arch.pmcr_n = new_n;
> > > > + else
> > > > + ret = -EINVAL;
> > > > + }
> > >
> > > Since we have set the default value of pmcr_n, if we want to set a new
> > > pmcr_n, shouldn't it be a different value?
> > >
> > > So how about change the checking to:
> > >
> > > if (likely(new_n <= kvm->arch.pmcr_n_limit)
> > > kvm->arch.pmcr_n = new_n;
> > > else
> > > ret = -EINVAL;
> > >
> > > what do you think?
> > >
> > Sorry, I guess I didn't fully understand your suggestion. Are you
> > saying that it's 'likely' that userspace would configure the correct
> > value?
> >
> It depends on how userspace use this api to limit the number of pmcr. I
> think what you mean in the code is that userspace need to set every vcpu's
> pmcr to the same value, so the `unlikely` here is right, only one vcpu can
> change the kvm->arch.pmcr.n, it saves the cpu cycles.
>
> What suggest above might be wrong. Since I think when userspace want to
> limit the number of pmcr, it may just set the new_n on one vcpu, since the
> kvm->arch.pmcr_n is a VM-local value, every vcpu can see it, so it's
> `likely` the (new_n <= kvm->arch.pmcr_n_limit), it can decrease one checking
> statement.
How about we just do away with branch hints in the first place? This is
_not_ a hot path.
--
Thanks,
Oliver
On Fri, Sep 15, 2023 at 1:36 PM Oliver Upton <[email protected]> wrote:
>
> On Tue, Aug 22, 2023 at 11:26:23AM +0800, Shaoqin Huang wrote:
>
> [...]
>
> > > > > +static int set_pmcr(struct kvm_vcpu *vcpu, const struct sys_reg_desc *r,
> > > > > + u64 val)
> > > > > +{
> > > > > + struct kvm *kvm = vcpu->kvm;
> > > > > + u64 new_n, mutable_mask;
> > > > > + int ret = 0;
> > > > > +
> > > > > + new_n = FIELD_GET(ARMV8_PMU_PMCR_N, val);
> > > > > +
> > > > > + mutex_lock(&kvm->arch.config_lock);
> > > > > + if (unlikely(new_n != kvm->arch.pmcr_n)) {
> > > > > + /*
> > > > > + * The vCPU can't have more counters than the PMU
> > > > > + * hardware implements.
> > > > > + */
> > > > > + if (new_n <= kvm->arch.pmcr_n_limit)
> > > > > + kvm->arch.pmcr_n = new_n;
> > > > > + else
> > > > > + ret = -EINVAL;
> > > > > + }
> > > >
> > > > Since we have set the default value of pmcr_n, if we want to set a new
> > > > pmcr_n, shouldn't it be a different value?
> > > >
> > > > So how about change the checking to:
> > > >
> > > > if (likely(new_n <= kvm->arch.pmcr_n_limit)
> > > > kvm->arch.pmcr_n = new_n;
> > > > else
> > > > ret = -EINVAL;
> > > >
> > > > what do you think?
> > > >
> > > Sorry, I guess I didn't fully understand your suggestion. Are you
> > > saying that it's 'likely' that userspace would configure the correct
> > > value?
> > >
> > It depends on how userspace use this api to limit the number of pmcr. I
> > think what you mean in the code is that userspace need to set every vcpu's
> > pmcr to the same value, so the `unlikely` here is right, only one vcpu can
> > change the kvm->arch.pmcr.n, it saves the cpu cycles.
> >
> > What suggest above might be wrong. Since I think when userspace want to
> > limit the number of pmcr, it may just set the new_n on one vcpu, since the
> > kvm->arch.pmcr_n is a VM-local value, every vcpu can see it, so it's
> > `likely` the (new_n <= kvm->arch.pmcr_n_limit), it can decrease one checking
> > statement.
>
> How about we just do away with branch hints in the first place? This is
> _not_ a hot path.
>
Sounds good to me.
Thank you.
Raghavendra
> --
> Thanks,
> Oliver