On 8/14/2025 6:41 PM, Peter Zijlstra wrote:
On Thu, Aug 14, 2025 at 03:10:46AM +0000, Li,Rongqing wrote:
On 8/13/2025 10:33 PM, Peter Zijlstra wrote:
On Wed, Aug 13, 2025 at 08:50:43AM +0800, Wangyang Guo wrote:
Binary exponential backoff is introduced. As try-lock attempt
increases, there is more likely that a larger number threads compete
for the same lock, so increase wait time in exponential.
You shouldn't be using virt_spin_lock() to begin with. That means
you've misconfigured your guest.
We have paravirt spinlocks for a reason.
We have tried PARAVIRT_SPINLOCKS, it can help to reduce the contention cycles,
but the throughput is not good. I think there are two factors:
1. the VM is not overcommit, each thread has its CPU resources to doing spin
wait.
If vm is not overcommit, guest should have KVM_HINTS_REALTIME, I think native qspinlock should be better
Could you try test this patch
https://patchwork.kernel.org/project/kvm/patch/20250722110005.4988-1-lirongqing@xxxxxxxxx/
Right, that's the knob.
Yes, that works.
By providing qemu-kvm with `-cpu host,kvm-hint-dedicated` option, it can
use mcs qspinlock and have better performance compared to virt_spin_lock().
But for non-overcommit VM, we may add `-overcommit cpu-pm=on` to let
guest to handle idle by itself and reduce the guest latency. Current
kernel will fallback to virt_spin_lock, even kvm-hint-dedicated is
provided. With Rongqing's patch, it can fix this problem.
BR
Wangyang
