Re: [PATCH][v2] x86/kvm: Prefer native qspinlock for dedicated vCPUs irrespective of PV_UNHALT

["Guo, Wangyang" <wangyang.guo@intel.com>]

On 7/22/2025 7:00 PM, lirongqing wrote:
> From: Li RongQing <lirongqing@baidu.com>
> 
> The commit b2798ba0b876 ("KVM: X86: Choose qspinlock when dedicated
> physical CPUs are available") states that when PV_DEDICATED=1
> (vCPU has dedicated pCPU), qspinlock should be preferred regardless of
> PV_UNHALT.  However, the current implementation doesn't reflect this: when
> PV_UNHALT=0, we still use virt_spin_lock() even with dedicated pCPUs.
> 
> This is suboptimal because:
> 1. Native qspinlocks should outperform virt_spin_lock() for dedicated
>     vCPUs irrespective of HALT exiting
> 2. virt_spin_lock() should only be preferred when vCPUs may be preempted
>     (non-dedicated case)
> 
> So reorder the PV spinlock checks to:
> 1. First handle dedicated pCPU case (disable virt_spin_lock_key)
> 2. Second check single CPU, and nopvspin configuration
> 3. Only then check PV_UNHALT support
> 
> This ensures we always use native qspinlock for dedicated vCPUs, delivering
> pretty performance gains at high contention levels.
> 
> Signed-off-by: Li RongQing <lirongqing@baidu.com>
> 
> diff with v1: rewrite the changelog
> 
>   arch/x86/kernel/kvm.c | 20 ++++++++++----------
>   1 file changed, 10 insertions(+), 10 deletions(-)
> 
> diff --git a/arch/x86/kernel/kvm.c b/arch/x86/kernel/kvm.c
> index 921c1c7..9cda79f 100644
> --- a/arch/x86/kernel/kvm.c
> +++ b/arch/x86/kernel/kvm.c
> @@ -1073,16 +1073,6 @@ static void kvm_wait(u8 *ptr, u8 val)
>   void __init kvm_spinlock_init(void)
>   {
>   	/*
> -	 * In case host doesn't support KVM_FEATURE_PV_UNHALT there is still an
> -	 * advantage of keeping virt_spin_lock_key enabled: virt_spin_lock() is
> -	 * preferred over native qspinlock when vCPU is preempted.
> -	 */
> -	if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT)) {
> -		pr_info("PV spinlocks disabled, no host support\n");
> -		return;
> -	}
> -
> -	/*
>   	 * Disable PV spinlocks and use native qspinlock when dedicated pCPUs
>   	 * are available.
>   	 */
> @@ -1101,6 +1091,16 @@ void __init kvm_spinlock_init(void)
>   		goto out;
>   	}
>   
> +	/*
> +	 * In case host doesn't support KVM_FEATURE_PV_UNHALT there is still an
> +	 * advantage of keeping virt_spin_lock_key enabled: virt_spin_lock() is
> +	 * preferred over native qspinlock when vCPU is preempted.
> +	 */
> +	if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT)) {
> +		pr_info("PV spinlocks disabled, no host support\n");
> +		return;
> +	}
> +
>   	pr_info("PV spinlocks enabled\n");
>   
>   	__pv_init_lock_hash();

For non-overcommit VM, we may add `-overcommit cpu-pm=on` options to 
qemu-kvm and let guest to handle idle by itself and reduce the latency. 
Current kernel will fallback to virt_spin_lock, even kvm-hint-dedicated 
is provided. With this patch, it can fix this problem and use mcs queue 
spinlock for better performance.

Tested-by: Wangyang Guo <wangyang.guo@intel.com>