Re: [PATCH] KVM: Dynamically scale directed-yield retries in kvm_vcpu_on_spin()

[Wei Wang]

On 4/20/26 10:31 AM, zhanghao wrote:
> kvm_vcpu_on_spin() currently uses a fixed retry budget (3) when
> kvm_vcpu_yield_to() returns a negative value. That fixed limit is
> often too small on large VMs, where candidate selection is more likely
> to race with scheduler state and transiently fail.
>
> Replace the constant retry count with a vCPU-count-based budget:
>
>    try = min(nr_vcpus - 1, 8);

Why is 8 the right value here?

Also, the probability of hitting these races depends more on host
overcommitment than on the VM’s vCPU count.

> This keeps retries bounded while allowing more opportunities to find a
> yieldable target as VM size grows. Small VMs keep near-identical
> behavior, while larger VMs get a better chance of making directed-yield
> progress under contention.

Increasing the retry count also adds overhead. Do you have data showing
that this could provide a measurable performance benefit?

> No functional change is intended beyond retry-budget tuning.
>
> Signed-off-by: Hao Zhang <zhanghao1@xxxxxxxxxx>
> ---
>   virt/kvm/kvm_main.c | 4 +++-
>   1 file changed, 3 insertions(+), 1 deletion(-)
>
> diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
> index 89489996fbc1..cbe53cea10a3 100644
> --- a/virt/kvm/kvm_main.c
> +++ b/virt/kvm/kvm_main.c
> @@ -3961,12 +3961,14 @@ void kvm_vcpu_on_spin(struct kvm_vcpu *me, bool yield_to_kernel_mode)
>   	int nr_vcpus, start, i, idx, yielded;
>   	struct kvm *kvm = me->kvm;
>   	struct kvm_vcpu *vcpu;
> -	int try = 3;
> +	int try = 0;
>   
>   	nr_vcpus = atomic_read(&kvm->online_vcpus);
>   	if (nr_vcpus < 2)
>   		return;
>   
> +	try = min(nr_vcpus - 1, 8);
> +
>   	/* Pairs with the smp_wmb() in kvm_vm_ioctl_create_vcpu(). */
>   	smp_rmb();
>   
>
> base-commit: c1f49dea2b8f335813d3b348fd39117fb8efb428