[v2,16/43] KVM: Don't redo ktime_get() when calculating halt-polling stop/deadline

Commit Message

Sean Christopherson Oct. 9, 2021, 2:12 a.m. UTC

Calculate the halt-polling "stop" time using "cur" instead of redoing
ktime_get().  In the happy case where hardware correctly predicts
do_halt_poll, "cur" is only a few cycles old.  And if the branch is
mispredicted, arguably that extra latency should count toward the
halt-polling time.

In all likelihood, the numbers involved are in the noise and either
approach is perfectly ok.

Reviewed-by: David Matlack <[e-mail]>
Signed-off-by: Sean Christopherson <[e-mail]>
---
 virt/kvm/kvm_main.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

Comments

Paolo Bonzini Oct. 25, 2021, 2:26 p.m. UTC | #1

On 09/10/21 04:12, Sean Christopherson wrote:
> Calculate the halt-polling "stop" time using "cur" instead of redoing
> ktime_get().  In the happy case where hardware correctly predicts
> do_halt_poll, "cur" is only a few cycles old.  And if the branch is
> mispredicted, arguably that extra latency should count toward the
> halt-polling time.
> 
> In all likelihood, the numbers involved are in the noise and either
> approach is perfectly ok.

Using "start" makes the change even more obvious, so:

     Calculate the halt-polling "stop" time using "start" instead of redoing
     ktime_get().  In practice, the numbers involved are in the noise (e.g.,
     in the happy case where hardware correctly predicts do_halt_poll and
     there are no interrupts, "start" is probably only a few cycles old)
     and either approach is perfectly ok.  But it's more precise to count
     any extra latency toward the halt-polling time.

Paolo

Maxim Levitsky Oct. 27, 2021, 2:35 p.m. UTC | #2

On Mon, 2021-10-25 at 16:26 +0200, Paolo Bonzini wrote:
> On 09/10/21 04:12, Sean Christopherson wrote:
> > Calculate the halt-polling "stop" time using "cur" instead of redoing
> > ktime_get().  In the happy case where hardware correctly predicts
> > do_halt_poll, "cur" is only a few cycles old.  And if the branch is
> > mispredicted, arguably that extra latency should count toward the
> > halt-polling time.
> > 
> > In all likelihood, the numbers involved are in the noise and either
> > approach is perfectly ok.
> 
> Using "start" makes the change even more obvious, so:
> 
>      Calculate the halt-polling "stop" time using "start" instead of redoing
>      ktime_get().  In practice, the numbers involved are in the noise (e.g.,
>      in the happy case where hardware correctly predicts do_halt_poll and
>      there are no interrupts, "start" is probably only a few cycles old)
>      and either approach is perfectly ok.  But it's more precise to count
>      any extra latency toward the halt-polling time.
> 
> Paolo
> 
Agreed.

Reviewed-by: Maxim Levitsky <[e-mail]>

Patch

diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
index a36ccdc93a72..481e8178b43d 100644
--- a/virt/kvm/kvm_main.c
+++ b/virt/kvm/kvm_main.c
@@ -3272,7 +3272,7 @@  void kvm_vcpu_halt(struct kvm_vcpu *vcpu)
 
 	start = cur = poll_end = ktime_get();
 	if (do_halt_poll) {
-		ktime_t stop = ktime_add_ns(ktime_get(), vcpu->halt_poll_ns);
+		ktime_t stop = ktime_add_ns(cur, vcpu->halt_poll_ns);
 
 		do {
 			/*