On 2026/09/22 15:54, Oliver Upton wrote:
On Sun, Sep 20, 2026 at 08:15:50PM +0900, Akihiko Odaki wrote:
+void kvm_pmu_apply_mdcr(struct kvm_vcpu *vcpu, u64 old, u64 val)
+{
+ u64 changed = old ^ val;
+
+ /*
+ * HPMN determines which counters HPMD and HLP apply to. Changes to
+ * these fields require new perf event filters and sample periods.
+ */
+ if (changed & (MDCR_EL2_HPMN | MDCR_EL2_HPMD | MDCR_EL2_HLP))
+ kvm_pmu_request_recreate(vcpu);
+ else if (changed & MDCR_EL2_HPME)
+ kvm_make_request(KVM_REQ_RELOAD_PMU, vcpu);
+}
+
Urgh... I'm not a fan of this. Can you just modify kvm_vcpu_reload_pmu()
to discard all backing perf events and recompute them only for enabled
PMCs? We set KVM_REQ_RELOAD_PMU pretty rarely.
__vcpu_rmw_sys_reg(vcpu, PMOVSSET_EL0, &=, mask);
__vcpu_rmw_sys_reg(vcpu, PMINTENSET_EL1, &=, mask);
diff --git a/arch/arm64/kvm/sys_regs.c b/arch/arm64/kvm/sys_regs.c
index 75624725adf1..7dd2cafce247 100644
--- a/arch/arm64/kvm/sys_regs.c
+++ b/arch/arm64/kvm/sys_regs.c
@@ -3114,17 +3114,22 @@ static bool access_mdcr(struct kvm_vcpu *vcpu,
}
__vcpu_assign_sys_reg(vcpu, MDCR_EL2, val);
-
- /*
- * Request a reload of the PMU to enable/disable the counters
- * affected by HPME.
- */
- if ((old ^ val) & MDCR_EL2_HPME)
- kvm_make_request(KVM_REQ_RELOAD_PMU, vcpu);
+ kvm_pmu_apply_mdcr(vcpu, old, val);
return true;
}
+static int set_mdcr(struct kvm_vcpu *vcpu, const struct sys_reg_desc *rd,
+ u64 val)
+{
+ u64 old = __vcpu_sys_reg(vcpu, MDCR_EL2);
+
+ __vcpu_assign_sys_reg(vcpu, MDCR_EL2, val);
+ kvm_pmu_apply_mdcr(vcpu, old, val);
+
+ return 0;
+}
+
Just keep the old ^ new test inline for access_mdcr() and set the
request unconditionally from set_mdcr(). In the usual case
KVM_REQ_RELOAD_PMU is already set on a vCPU that userspace is restoring.
Hi Oliver,
Thanks for the feedback. I'll make both changes for v2.
I originally tried to preserve existing behavior and keep things
consistent, but I agree that simplicity is much better here given how
rarely KVM_REQ_RELOAD_PMU is set.
Regards,
Akihiko Odaki
