I noticed that srcu_read_lock/unlock both have a memory barrier,
so just by moving srcu_read_unlock earlier we can get rid of
one call to smp_mb() using smp_mb__after_srcu_read_unlock instead.
Unsurprisingly, the gain is small but measureable using the unit test
microbenchmark:
before
vmcall in the ballpark of 1410 cycles
after
vmcall in the ballpark of 1360 cycles
arch/x86/kvm/x86.c | 7 ++++---
1 file changed, 4 insertions(+), 3 deletions(-)
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index e5ca72a..d609dce 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -5945,10 +5945,12 @@ static int vcpu_enter_guest(struct kvm_vcpu *vcpu)
vcpu->mode = IN_GUEST_MODE;
+ srcu_read_unlock(&vcpu->kvm->srcu, vcpu->srcu_idx);
+
/* We should set ->mode before check ->requests,
* see the comment in make_all_cpus_request.
*/
- smp_mb();
+ smp_mb__after_srcu_read_unlock();
local_irq_disable();
@@ -5958,12 +5960,11 @@ static int vcpu_enter_guest(struct kvm_vcpu *vcpu)
smp_wmb();
local_irq_enable();
preempt_enable();
+ vcpu->srcu_idx = srcu_read_lock(&vcpu->kvm->srcu);
r = 1;
goto cancel_injection;
}
- srcu_read_unlock(&vcpu->kvm->srcu, vcpu->srcu_idx);
-
if (req_immediate_exit)
smp_send_reschedule(vcpu->cpu);