Skip to contentSkip to conversation

KVM: x86: Set PF_VCPU when processing IRQs to fix tick-based accounting

RFC PATCH In review Sean Christopherson 2 participants 2 messages Feb 6, 2021 – Feb 8, 2021

Subsystems (3)
lore
Share & tools
lore

Sean Christopherson

author maintainer
#0 lore raw

RFC PATCH 1/1 KVM: x86: Set PF_VCPU when processing IRQs to fix tick-based accounting

Temporarily set PF_VCPU while processing IRQ VM-Exits so that a tick IRQ accounts the time slice to the guest. Tick-based accounting of guest time is currently broken as PF_VCPU is only set during the relatively short VM-Enter sequence, which runs entirely with IRQs disabled, and IRQs that occur in the guest are processed well after PF_VCPU is cleared.

Keep PF_VCPU set across both VMX's processing of already-acked IRQs (handle_exit_irqoff()) and the explicit IRQ window (SVM's processing, plus ticks that occur immediately after VM-Exit on both VMX and SVM).

Fixes 87fa7f3e98a1 ("x86/kvm: Move context tracking where it belongs")
Cc Peter Zijlstra <[email protected]>
Cc Thomas Gleixner <[email protected]>

---

This is quite obnoxious, hence the RFC, but I can't think of a clever, less ugly way to fix the accounting.

 arch/x86/kvm/x86.c | 8 ++++++++
 1 file changed, 8 insertions(+)
arch/x86/kvm/x86.c +8 -0
index d9f931c63293..6ddf341cd755 100644
@@ -9118,6 +9118,13 @@ static int vcpu_enter_guest(struct kvm_vcpu *vcpu)
 	vcpu->mode = OUTSIDE_GUEST_MODE;
 	smp_wmb();
 
+	/*
+	 * Temporarily pretend this task is running a vCPU when potentially
+	 * processing an IRQ exit, including the below opening of an IRQ
+	 * window.  Tick-based accounting of guest time relies on PF_VCPU
+	 * being set when the tick IRQ handler runs.
+	 */
+	current->flags |= PF_VCPU;
 	static_call(kvm_x86_handle_exit_irqoff)(vcpu);
 
 	/*
@@ -9132,6 +9139,7 @@ static int vcpu_enter_guest(struct kvm_vcpu *vcpu)
 	++vcpu->stat.exits;
 	local_irq_disable();
 	kvm_after_interrupt(vcpu);
+	current->flags &= ~PF_VCPU;
 
 	if (lapic_in_kernel(vcpu)) {
 		s64 delta = vcpu->arch.apic->lapic_timer.advance_expire_delta;
Signature · 1 line
2.30.0.478.g8a0d178c01-goog
#1 lore raw
Quoted 28 lines · Sean Christopherson

Temporarily set PF_VCPU while processing IRQ VM-Exits so that a tick IRQ accounts the time slice to the guest. Tick-based accounting of guest time is currently broken as PF_VCPU is only set during the relatively short VM-Enter sequence, which runs entirely with IRQs disabled, and IRQs that occur in the guest are processed well after PF_VCPU is cleared.

Keep PF_VCPU set across both VMX's processing of already-acked IRQs (handle_exit_irqoff()) and the explicit IRQ window (SVM's processing, plus ticks that occur immediately after VM-Exit on both VMX and SVM).

Fixes 87fa7f3e98a1 ("x86/kvm: Move context tracking where it belongs")
Cc Peter Zijlstra <[email protected]>
Cc Thomas Gleixner <[email protected]>

---

This is quite obnoxious, hence the RFC, but I can't think of a clever, less ugly way to fix the accounting.

 arch/x86/kvm/x86.c | 8 ++++++++
 1 file changed, 8 insertions(+)
arch/x86/kvm/x86.c +0 -0
index d9f931c63293..6ddf341cd755 100644

arch/x86/kvm/x86.c static int vcpu_enter_guest(struct kvm_vcpu *vcpu) · line 9127

Quoted 10 lines
arch/x86/kvm/x86.c +7 -0
@@ -9118,6 +9118,13 @@ static int vcpu_enter_guest(struct kvm_vcpu *vcpu)
 	vcpu->mode = OUTSIDE_GUEST_MODE;
 	smp_wmb();
 
+	/*
+	 * Temporarily pretend this task is running a vCPU when potentially
+	 * processing an IRQ exit, including the below opening of an IRQ
+	 * window.  Tick-based accounting of guest time relies on PF_VCPU
+	 * being set when the tick IRQ handler runs.
+	 */
+	current->flags |= PF_VCPU;

Should we do it only when !vtime_accounting_enabled_cpu() maybe?

Quoted 11 lines
arch/x86/kvm/x86.c +1 -0
 	static_call(kvm_x86_handle_exit_irqoff)(vcpu);
 
 	/*
@@ -9132,6 +9139,7 @@ static int vcpu_enter_guest(struct kvm_vcpu *vcpu)
 	++vcpu->stat.exits;
 	local_irq_disable();
 	kvm_after_interrupt(vcpu);
+	current->flags &= ~PF_VCPU;
 
 	if (lapic_in_kernel(vcpu)) {
 		s64 delta = vcpu->arch.apic->lapic_timer.advance_expire_delta;
Signature · 1 line
Vitaly