kvm: x86: avoid atomic operations on APICv vmentry
On some benchmarks (e.g. netperf with ioeventfd disabled), APICv
posted interrupts turn out to be slower than interrupt injection via
KVM_REQ_EVENT.
This patch optimizes a bit the IRR update, avoiding expensive atomic
operations in the common case where PI.ON=0 at vmentry or the PIR vector
is mostly zero. This saves at least 20 cycles (1%) per vmexit, as
measured by kvm-unit-tests' inl_from_qemu test (20 runs):
| enable_apicv=1 | enable_apicv=0
| mean stdev | mean stdev
----------|-----------------|------------------
before | 5826 32.65 | 5765 47.09
after | 5809 43.42 | 5777 77.02
Of course, any change in the right column is just placebo effect. :)
The savings are bigger if interrupts are frequent.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
diff --git a/arch/x86/kvm/vmx.c b/arch/x86/kvm/vmx.c
index dcb840e..a91a5b0 100644
--- a/arch/x86/kvm/vmx.c
+++ b/arch/x86/kvm/vmx.c
@@ -520,6 +520,12 @@ static inline void pi_set_sn(struct pi_desc *pi_desc)
(unsigned long *)&pi_desc->control);
}
+static inline void pi_clear_on(struct pi_desc *pi_desc)
+{
+ clear_bit(POSTED_INTR_ON,
+ (unsigned long *)&pi_desc->control);
+}
+
static inline int pi_test_on(struct pi_desc *pi_desc)
{
return test_bit(POSTED_INTR_ON,
@@ -4780,9 +4786,15 @@ static void vmx_sync_pir_to_irr(struct kvm_vcpu *vcpu)
{
struct vcpu_vmx *vmx = to_vmx(vcpu);
- if (!pi_test_and_clear_on(&vmx->pi_desc))
+ if (!pi_test_on(&vmx->pi_desc))
return;
+ pi_clear_on(&vmx->pi_desc);
+ /*
+ * IOMMU can write to PIR.ON, so the barrier matters even on UP.
+ * But on x86 this is just a compiler barrier anyway.
+ */
+ smp_mb__after_atomic();
kvm_apic_update_irr(vcpu, vmx->pi_desc.pir);
}