KVM: X86: Narrow down the IPI fastpath to single target IPI

The original single target IPI fastpath patch forgot to filter the
ICR destination shorthand field. Multicast IPI is not suitable for
this feature since wakeup the multiple sleeping vCPUs will extend
the interrupt disabled time, it especially worse in the over-subscribe
and VM has a little bit more vCPUs scenario. Let's narrow it down to
single target IPI.

Two VMs, each is 76 vCPUs, one running 'ebizzy -M', the other
running cyclictest on all vCPUs, w/ this patch, the avg score
of cyclictest can improve more than 5%. (pv tlb, pv ipi, pv
sched yield are disabled during testing to avoid the disturb).

Signed-off-by: Wanpeng Li <wanpengli@tencent.com>
Message-Id: <1585189202-1708-3-git-send-email-wanpengli@tencent.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index d65ff2008cf1da2a6ccc182bcdda81c0f02ab4cc..cf95c36cb4f4aab2b5c0628460b8c41718ef5a01 100644 (file)
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -1554,7 +1554,10 @@ EXPORT_SYMBOL_GPL(kvm_emulate_wrmsr);
  */
 static int handle_fastpath_set_x2apic_icr_irqoff(struct kvm_vcpu *vcpu, u64 data)
 {
-       if (lapic_in_kernel(vcpu) && apic_x2apic_mode(vcpu->arch.apic) &&
+       if (!lapic_in_kernel(vcpu) || !apic_x2apic_mode(vcpu->arch.apic))
+               return 1;
+
+       if (((data & APIC_SHORT_MASK) == APIC_DEST_NOSHORT) &&
                ((data & APIC_DEST_MASK) == APIC_DEST_PHYSICAL) &&
                ((data & APIC_MODE_MASK) == APIC_DM_FIXED)) {