Skip to contentSkip to conversation

KVM: VMX: optimize APIC EOI

PATCH Marcelo Tosatti 2 participants 2 messages Jun 14, 2010 – Jun 15, 2010

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

Use information provided in exit_qualification to shortcut EOI path.

Reduces EOI latency from 4k to 2k cycles on Nehalem.

arch/x86/kvm/lapic.c +6 -0
Index: kvm/arch/x86/kvm/lapic.c
===================================================================
@@ -1279,3 +1279,9 @@ int kvm_hv_vapic_msr_read(struct kvm_vcp
 
 	return 0;
 }
+
+void kvm_lapic_set_eoi(struct kvm_vcpu *vcpu)
+{
+	apic_set_eoi(vcpu->arch.apic);
+}
+EXPORT_SYMBOL_GPL(kvm_lapic_set_eoi);
arch/x86/kvm/lapic.h +2 -0
Index: kvm/arch/x86/kvm/lapic.h
===================================================================
@@ -52,6 +52,8 @@ int kvm_x2apic_msr_read(struct kvm_vcpu 
 int kvm_hv_vapic_msr_write(struct kvm_vcpu *vcpu, u32 msr, u64 data);
 int kvm_hv_vapic_msr_read(struct kvm_vcpu *vcpu, u32 msr, u64 *data);
 
+void kvm_lapic_set_eoi(struct kvm_vcpu *vcpu);
+
 static inline bool kvm_hv_vapic_assist_page_enabled(struct kvm_vcpu *vcpu)
 {
 	return vcpu->arch.hv_vapic & HV_X64_MSR_APIC_ASSIST_PAGE_ENABLE;
arch/x86/kvm/vmx.c +11 -0
Index: kvm/arch/x86/kvm/vmx.c
===================================================================
@@ -3395,6 +3395,17 @@ static int handle_wbinvd(struct kvm_vcpu
 
 static int handle_apic_access(struct kvm_vcpu *vcpu)
 {
+	unsigned long exit_qualification = vmcs_readl(EXIT_QUALIFICATION);
+	int access_type, offset;
+
+	access_type = (exit_qualification >> 12) & 0xf;
+	offset = exit_qualification & 0xfff;
+	if (access_type == 1 && offset == APIC_EOI) {
+		kvm_lapic_set_eoi(vcpu);
+		skip_emulated_instruction(vcpu);
+		return 1;
+	}
+
 	return emulate_instruction(vcpu, 0, 0, 0) == EMULATE_DONE;
 }
 
#1 lore raw
Quoted 5 lines · Marcelo Tosatti

Use information provided in exit_qualification to shortcut EOI path.

Reduces EOI latency from 4k to 2k cycles on Nehalem.

This already came up once - IIRC from Eddie Dong.

Quoted 8 lines
  static int handle_apic_access(struct kvm_vcpu *vcpu)
  {
arch/x86/kvm/vmx.c +3 -0
+	unsigned long exit_qualification = vmcs_readl(EXIT_QUALIFICATION);
+	int access_type, offset;
+

arch/x86/kvm/vmx.c static int handle_wbinvd(struct kvm_vcpu · line 3410

Quoted 10 lines
arch/x86/kvm/vmx.c +8 -0
+	access_type = (exit_qualification>>  12)&  0xf;
+	offset = exit_qualification&  0xfff;
+	if (access_type == 1&&  offset == APIC_EOI) {
+		kvm_lapic_set_eoi(vcpu);
+		skip_emulated_instruction(vcpu);
+		return 1;
+	}
+
  	return emulate_instruction(vcpu, 0, 0, 0) == EMULATE_DONE;
  }

This fails if the instruction is STOS, MOVS, or a RMW instruction as it doesn't update registers and flags.

We could try to do tricks to detect this, but we already have x2apic for Linux and the corresponding Hyper-V extension, which improve a lot of other stuff, at least for newer guests.

\

Signature · 2 lines
I have a truly marvellous patch that fixes the bug which this
signature is too narrow to contain.