[2/3] KVM: X86: Bail out of direct yield in case of undercomitted scenarios

Commit Message

Wanpeng Li May 8, 2021, 9:31 a.m. UTC

From: Wanpeng Li <[e-mail]>

In case of undercomitted scenarios, vCPU can get scheduling easily, 
kvm_vcpu_yield_to adds extra overhead, we can observe a lot of race 
between vcpu->ready is true and yield fails due to p->state is 
TASK_RUNNING. Let's bail out is such scenarios by checking the length 
of current cpu runqueue.

Signed-off-by: Wanpeng Li <[e-mail]>
---
 arch/x86/kvm/x86.c | 3 +++
 1 file changed, 3 insertions(+)

Comments

Sean Christopherson May 11, 2021, 9:44 p.m. UTC | #1

On Sat, May 08, 2021, Wanpeng Li wrote:
> From: Wanpeng Li <[e-mail]>
> 
> In case of undercomitted scenarios, vCPU can get scheduling easily, 
> kvm_vcpu_yield_to adds extra overhead, we can observe a lot of race 
> between vcpu->ready is true and yield fails due to p->state is 
> TASK_RUNNING. Let's bail out is such scenarios by checking the length 
> of current cpu runqueue.
> 
> Signed-off-by: Wanpeng Li <[e-mail]>
> ---
>  arch/x86/kvm/x86.c | 3 +++
>  1 file changed, 3 insertions(+)
> 
> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> index 5bd550e..c0244a6 100644
> --- a/arch/x86/kvm/x86.c
> +++ b/arch/x86/kvm/x86.c
> @@ -8358,6 +8358,9 @@ static void kvm_sched_yield(struct kvm_vcpu *vcpu, unsigned long dest_id)
>  	struct kvm_vcpu *target = NULL;
>  	struct kvm_apic_map *map;
>  
> +	if (single_task_running())
> +		goto no_yield;
> +

Hmm, could we push the result of kvm_sched_yield() down into the guest?
Currently the guest bails after the first attempt, which is perfect for this
scenario, but it seems like it would make sense to keep trying to yield if there
are multiple preempted vCPUs and the "problem" was with the target.  E.g.

	/*
	 * Make sure other vCPUs get a chance to run if they need to.  Yield at
	 * most once, and stop trying to yield if the VMM says yielding isn't
	 * going to happen.
	 */
	for_each_cpu(cpu, mask) {
		if (vcpu_is_preempted(cpu)) {
			r = kvm_hypercall1(KVM_HC_SCHED_YIELD,
					   per_cpu(x86_cpu_to_apicid, cpu));
			if (r != -EBUSY)
				break;
		}
	}


Unrelated to this patch, but it's the first time I've really looked at the guest
side of directed yield...

Wouldn't it also make sense for the guest side to hook .send_call_func_single_ipi?

>  	vcpu->stat.directed_yield_attempted++;

Shouldn't directed_yield_attempted be incremented in this case?  It doesn't seem
fundamentally different than the case where the target was scheduled in between
the guest's check and the host's processing of the yield request.  In both
instances, the guest did indeed attempt to yield.

>  	rcu_read_lock();
> -- 
> 2.7.4
>

Wanpeng Li May 12, 2021, 2:43 a.m. UTC | #2

On Wed, 12 May 2021 at 05:44, Sean Christopherson <[e-mail]> wrote:
>
> On Sat, May 08, 2021, Wanpeng Li wrote:
> > From: Wanpeng Li <[e-mail]>
> >
> > In case of undercomitted scenarios, vCPU can get scheduling easily,
> > kvm_vcpu_yield_to adds extra overhead, we can observe a lot of race
> > between vcpu->ready is true and yield fails due to p->state is
> > TASK_RUNNING. Let's bail out is such scenarios by checking the length
> > of current cpu runqueue.
> >
> > Signed-off-by: Wanpeng Li <[e-mail]>
> > ---
> >  arch/x86/kvm/x86.c | 3 +++
> >  1 file changed, 3 insertions(+)
> >
> > diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> > index 5bd550e..c0244a6 100644
> > --- a/arch/x86/kvm/x86.c
> > +++ b/arch/x86/kvm/x86.c
> > @@ -8358,6 +8358,9 @@ static void kvm_sched_yield(struct kvm_vcpu *vcpu, unsigned long dest_id)
> >       struct kvm_vcpu *target = NULL;
> >       struct kvm_apic_map *map;
> >
> > +     if (single_task_running())
> > +             goto no_yield;
> > +
>
> Hmm, could we push the result of kvm_sched_yield() down into the guest?
> Currently the guest bails after the first attempt, which is perfect for this
> scenario, but it seems like it would make sense to keep trying to yield if there
> are multiple preempted vCPUs and

It can have a race in case of sustain yield if there are multiple
preempted vCPUs , the vCPU which you intend to yield may have already
completed to handle IPI and be preempted now when the yielded sender
is scheduled again and checks the next preempted candidate.

> the "problem" was with the target.  E.g.

At the beginning of kvm_sched_yield() we can just get the run queue
length of the source, it can be treated as a hint of under-committed
instead of guarantee of accuracy.

>
>         /*
>          * Make sure other vCPUs get a chance to run if they need to.  Yield at
>          * most once, and stop trying to yield if the VMM says yielding isn't
>          * going to happen.
>          */
>         for_each_cpu(cpu, mask) {
>                 if (vcpu_is_preempted(cpu)) {
>                         r = kvm_hypercall1(KVM_HC_SCHED_YIELD,
>                                            per_cpu(x86_cpu_to_apicid, cpu));
>                         if (r != -EBUSY)
>                                 break;
>                 }
>         }
>
>
> Unrelated to this patch, but it's the first time I've really looked at the guest
> side of directed yield...
>
> Wouldn't it also make sense for the guest side to hook .send_call_func_single_ipi?

reschedule ipi is called by .smp_send_reschedule hook, there are a lot
of researches intend to accelerate idle vCPU reactivation, my original
attemption is to boost synchronization primitive, I believe we need a
lot of benchmarkings to consider inter-VM fairness and performance
benefit for  hooks .send_call_func_single_ipi and
.smp_send_reschedule.

>
> >       vcpu->stat.directed_yield_attempted++;
>
> Shouldn't directed_yield_attempted be incremented in this case?  It doesn't seem
> fundamentally different than the case where the target was scheduled in between
> the guest's check and the host's processing of the yield request.  In both
> instances, the guest did indeed attempt to yield.

Yes, it should be treated as attempted, I move it above the counting
because this patch helps improve successful ratio in under-committed
scenarios and easily shows me how much failure ratio leaves over. I
can move it after the counting in the next version.

    Wanpeng

Sean Christopherson May 12, 2021, 4:59 p.m. UTC | #3

On Wed, May 12, 2021, Wanpeng Li wrote:
> On Wed, 12 May 2021 at 05:44, Sean Christopherson <[e-mail]> wrote:
> >
> > On Sat, May 08, 2021, Wanpeng Li wrote:
> > > From: Wanpeng Li <[e-mail]>
> > >
> > > In case of undercomitted scenarios, vCPU can get scheduling easily,
> > > kvm_vcpu_yield_to adds extra overhead, we can observe a lot of race
> > > between vcpu->ready is true and yield fails due to p->state is
> > > TASK_RUNNING. Let's bail out is such scenarios by checking the length
> > > of current cpu runqueue.
> > >
> > > Signed-off-by: Wanpeng Li <[e-mail]>
> > > ---
> > >  arch/x86/kvm/x86.c | 3 +++
> > >  1 file changed, 3 insertions(+)
> > >
> > > diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> > > index 5bd550e..c0244a6 100644
> > > --- a/arch/x86/kvm/x86.c
> > > +++ b/arch/x86/kvm/x86.c
> > > @@ -8358,6 +8358,9 @@ static void kvm_sched_yield(struct kvm_vcpu *vcpu, unsigned long dest_id)
> > >       struct kvm_vcpu *target = NULL;
> > >       struct kvm_apic_map *map;
> > >
> > > +     if (single_task_running())
> > > +             goto no_yield;
> > > +
> >
> > Hmm, could we push the result of kvm_sched_yield() down into the guest?
> > Currently the guest bails after the first attempt, which is perfect for this
> > scenario, but it seems like it would make sense to keep trying to yield if there
> > are multiple preempted vCPUs and
> 
> It can have a race in case of sustain yield if there are multiple
> preempted vCPUs , the vCPU which you intend to yield may have already
> completed to handle IPI and be preempted now when the yielded sender
> is scheduled again and checks the next preempted candidate.

Ah, right, don't want to penalize the happy case.

> > Unrelated to this patch, but it's the first time I've really looked at the guest
> > side of directed yield...
> >
> > Wouldn't it also make sense for the guest side to hook .send_call_func_single_ipi?
> 
> reschedule ipi is called by .smp_send_reschedule hook, there are a lot
> of researches intend to accelerate idle vCPU reactivation, my original
> attemption is to boost synchronization primitive, I believe we need a
> lot of benchmarkings to consider inter-VM fairness and performance
> benefit for  hooks .send_call_func_single_ipi and
> .smp_send_reschedule.

I was thinking of the 2 vCPU case.  If the VM has 2 vCPUs, then this

	/*
	 * Choose the most efficient way to send an IPI. Note that the
	 * number of CPUs might be zero due to concurrent changes to the
	 * provided mask.
	 */
	if (nr_cpus == 1)
		send_call_function_single_ipi(last_cpu);
	else if (likely(nr_cpus > 1))
		arch_send_call_function_ipi_mask(cfd->cpumask_ipi);

means .send_call_func_single_ipi() will always be used to send an IPI to the
other vCPU, and thus 2 vCPU VMs will never utilize PV yield.

Patch

diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index 5bd550e..c0244a6 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -8358,6 +8358,9 @@  static void kvm_sched_yield(struct kvm_vcpu *vcpu, unsigned long dest_id)
 	struct kvm_vcpu *target = NULL;
 	struct kvm_apic_map *map;
 
+	if (single_task_running())
+		goto no_yield;
+
 	vcpu->stat.directed_yield_attempted++;
 
 	rcu_read_lock();