Skip to contentSkip to conversation

KVM: s390: pending interrupts are unlikely

PATCH In review Michael Mueller 3 participants 3 messages Mar 13, 2020

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

PATCH 1/1 KVM: s390: pending interrupts are unlikely

A statistical analysis shows that in most cases when deliverable_irqs() is called, no interrupts are pending. (see: early exit ratio)

The data was sampled during an iperf3 run over virtio_net between one guest and the host.

deliverable_irqs()

        called = 3145123
           by kvm_s390_vcpu_has_irq() = 3005581 (95.56%)
              by kvm_arch_vcpu_runnable() = 3005578 (95.56%)
                 by kvm_s390_handle_wait() = 1219331 (38.76%)
                 by kvm_vcpu_check_block() = 2943565 (93.59%)
                    by kvm_cpu_block(1) = 2826431 (89.86%)
                    by kvm_cpu_block(2) = 117136 (3.72%)
                 by kvm_arch_dy_runnable() = 0 (0%)
              by kvm_arch_setup_async_pf() = 0 (0%)
              by handle_stop() = 0 (0%)
           by kvm_s390_deliver_pending_interrupt() = 139542 (4.43%)
              irqs_delivered = (0:15917 1:61810 2:1 3:0 4:0 x:0)
              irqs_pending = (0:15917 1:61722 2:86 3:1 4:0 x:0)
    early exit = 3021787 (96.07%)

pending irqs = 123237 (3.91%)

---
 arch/s390/kvm/interrupt.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
arch/s390/kvm/interrupt.c +1 -1
index 028167d6eacd..c34d62b4209e 100644
@@ -369,7 +369,7 @@ static unsigned long deliverable_irqs(struct kvm_vcpu *vcpu)
 	unsigned long active_mask;
 
 	active_mask = pending_irqs(vcpu);
-	if (!active_mask)
+	if (likely(!active_mask))
 		return 0;
 
 	if (psw_extint_disabled(vcpu))
Signature · 1 line
2.17.0
#1 lore raw
Quoted 34 lines · Michael Mueller

A statistical analysis shows that in most cases when deliverable_irqs() is called, no interrupts are pending. (see: early exit ratio)

The data was sampled during an iperf3 run over virtio_net between one guest and the host.

deliverable_irqs()

        called = 3145123
           by kvm_s390_vcpu_has_irq() = 3005581 (95.56%)
              by kvm_arch_vcpu_runnable() = 3005578 (95.56%)
                 by kvm_s390_handle_wait() = 1219331 (38.76%)
                 by kvm_vcpu_check_block() = 2943565 (93.59%)
                    by kvm_cpu_block(1) = 2826431 (89.86%)
                    by kvm_cpu_block(2) = 117136 (3.72%)
                 by kvm_arch_dy_runnable() = 0 (0%)
              by kvm_arch_setup_async_pf() = 0 (0%)
              by handle_stop() = 0 (0%)
           by kvm_s390_deliver_pending_interrupt() = 139542 (4.43%)
              irqs_delivered = (0:15917 1:61810 2:1 3:0 4:0 x:0)
              irqs_pending = (0:15917 1:61722 2:86 3:1 4:0 x:0)
    early exit = 3021787 (96.07%)

pending irqs = 123237 (3.91%)

---
 arch/s390/kvm/interrupt.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
arch/s390/kvm/interrupt.c +0 -0
index 028167d6eacd..c34d62b4209e 100644

arch/s390/kvm/interrupt.c static unsigned long deliverable_irqs(struct kvm_vcpu *vcpu) · line 375

Quoted 8 lines
arch/s390/kvm/interrupt.c +1 -1
@@ -369,7 +369,7 @@ static unsigned long deliverable_irqs(struct kvm_vcpu *vcpu)
 	unsigned long active_mask;
 
 	active_mask = pending_irqs(vcpu);
-	if (!active_mask)
+	if (likely(!active_mask))
 		return 0;
 
 	if (psw_extint_disabled(vcpu))

Is this change even observable in practice? Usually, we do have some performance numbers backing such micro optimizations. But I guess it will be fairly hard to get some meaning full numbers backing this ...

Signature · 3 lines
Thanks,

David / dhildenb
Acked-by on v1 1/1
#2 lore raw
In reply to Michael Mueller
Quoted 33 lines · Michael Mueller

A statistical analysis shows that in most cases when deliverable_irqs() is called, no interrupts are pending. (see: early exit ratio)

The data was sampled during an iperf3 run over virtio_net between one guest and the host.

deliverable_irqs()

        called = 3145123
           by kvm_s390_vcpu_has_irq() = 3005581 (95.56%)
              by kvm_arch_vcpu_runnable() = 3005578 (95.56%)
                 by kvm_s390_handle_wait() = 1219331 (38.76%)
                 by kvm_vcpu_check_block() = 2943565 (93.59%)
                    by kvm_cpu_block(1) = 2826431 (89.86%)
                    by kvm_cpu_block(2) = 117136 (3.72%)
                 by kvm_arch_dy_runnable() = 0 (0%)
              by kvm_arch_setup_async_pf() = 0 (0%)
              by handle_stop() = 0 (0%)
           by kvm_s390_deliver_pending_interrupt() = 139542 (4.43%)
              irqs_delivered = (0:15917 1:61810 2:1 3:0 4:0 x:0)
              irqs_pending = (0:15917 1:61722 2:86 3:1 4:0 x:0)
    early exit = 3021787 (96.07%)

pending irqs = 123237 (3.91%)

---
 arch/s390/kvm/interrupt.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
arch/s390/kvm/interrupt.c +0 -0
index 028167d6eacd..c34d62b4209e 100644

arch/s390/kvm/interrupt.c static unsigned long deliverable_irqs(struct kvm_vcpu *vcpu) · line 372

Quoted 5 lines
arch/s390/kvm/interrupt.c +1 -1
@@ -369,7 +369,7 @@ static unsigned long deliverable_irqs(struct kvm_vcpu *vcpu)
 	unsigned long active_mask;
 
 	active_mask = pending_irqs(vcpu);
-	if (!active_mask)
+	if (likely(!active_mask))

I have never been a fan of the likely/unlikely name. Instead this should have been fastpath/slowpath. And with likely actually meaning fastpath this patch makes sense.

I guess it does not change the performance as the branch predictor is doing its job most of the time. So I assume it would only be visible when there is contention in the host.

Tentatively

will have a look at the code.