On Tue, 20 Apr 2021 at 15:23, Paolo Bonzini <pbonzini@xxxxxxxxxx> wrote:
>
> On 20/04/21 08:08, Wanpeng Li wrote:
> > On Tue, 20 Apr 2021 at 14:02, Wanpeng Li <kernellwp@xxxxxxxxx> wrote:
> >>
> >> On Tue, 20 Apr 2021 at 00:59, Paolo Bonzini <pbonzini@xxxxxxxxxx> wrote:
> >>>
> >>> On 19/04/21 18:32, Sean Christopherson wrote:
> >>>> If false positives are a big concern, what about adding another pass to the loop
> >>>> and only yielding to usermode vCPUs with interrupts in the second full pass?
> >>>> I.e. give vCPUs that are already in kernel mode priority, and only yield to
> >>>> handle an interrupt if there are no vCPUs in kernel mode.
> >>>>
> >>>> kvm_arch_dy_runnable() pulls in pv_unhalted, which seems like a good thing.
> >>>
> >>> pv_unhalted won't help if you're waiting for a kernel spinlock though,
> >>> would it? Doing two passes (or looking for a "best" candidate that
> >>> prefers kernel mode vCPUs to user mode vCPUs waiting for an interrupt)
> >>> seems like the best choice overall.
> >>
> >> How about something like this:
>
> I was thinking of something simpler:
>
> diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
> index 9b8e30dd5b9b..455c648f9adc 100644
> --- a/virt/kvm/kvm_main.c
> +++ b/virt/kvm/kvm_main.c
> @@ -3198,10 +3198,9 @@ void kvm_vcpu_on_spin(struct kvm_vcpu *me, bool yield_to_kernel_mode)
> {
> struct kvm *kvm = me->kvm;
> struct kvm_vcpu *vcpu;
> - int last_boosted_vcpu = me->kvm->last_boosted_vcpu;
> int yielded = 0;
> int try = 3;
> - int pass;
> + int pass, num_passes = 1;
> int i;
>
> kvm_vcpu_set_in_spin_loop(me, true);
> @@ -3212,13 +3211,14 @@ void kvm_vcpu_on_spin(struct kvm_vcpu *me, bool yield_to_kernel_mode)
> * VCPU is holding the lock that we need and will release it.
> * We approximate round-robin by starting at the last boosted VCPU.
> */
> - for (pass = 0; pass < 2 && !yielded && try; pass++) {
> - kvm_for_each_vcpu(i, vcpu, kvm) {
> - if (!pass && i <= last_boosted_vcpu) {
> - i = last_boosted_vcpu;
> - continue;
> - } else if (pass && i > last_boosted_vcpu)
> - break;
> + for (pass = 0; pass < num_passes; pass++) {
> + int idx = me->kvm->last_boosted_vcpu;
> + int n = atomic_read(&kvm->online_vcpus);
> + for (i = 0; i < n; i++, idx++) {
> + if (idx == n)
> + idx = 0;
> +
> + vcpu = kvm_get_vcpu(kvm, idx);
> if (!READ_ONCE(vcpu->ready))
> continue;
> if (vcpu == me)
> @@ -3226,23 +3226,36 @@ void kvm_vcpu_on_spin(struct kvm_vcpu *me, bool yield_to_kernel_mode)
> if (rcuwait_active(&vcpu->wait) &&
> !vcpu_dy_runnable(vcpu))
> continue;
> - if (READ_ONCE(vcpu->preempted) && yield_to_kernel_mode &&
> - !kvm_arch_vcpu_in_kernel(vcpu))
> - continue;
> if (!kvm_vcpu_eligible_for_directed_yield(vcpu))
> continue;
>
> + if (READ_ONCE(vcpu->preempted) && yield_to_kernel_mode &&
> + !kvm_arch_vcpu_in_kernel(vcpu)) {
> + /*
> + * A vCPU running in userspace can get to kernel mode via
> + * an interrupt. That's a worse choice than a CPU already
> + * in kernel mode so only do it on a second pass.
> + */
> + if (!vcpu_dy_runnable(vcpu))
> + continue;
> + if (pass == 0) {
> + num_passes = 2;
> + continue;
> + }
> + }
> +
> yielded = kvm_vcpu_yield_to(vcpu);
> if (yielded > 0) {
> kvm->last_boosted_vcpu = i;
> - break;
> + goto done;
> } else if (yielded < 0) {
> try--;
> if (!try)
> - break;
> + goto done;
> }
> }
> }
> +done:
We just tested the above post against 96 vCPUs VM in an over-subscribe
scenario, the score of pbzip2 fluctuated drastically. Sometimes it is
worse than vanilla, but the average improvement is around 2.2%. The
new version of my post is around 9.3%，the origial posted patch is
around 10% which is totally as expected since now both IPI receivers
in user-mode and lock-waiters are second class citizens. Big VM
increases the probability multiple vCPUs may enter PLE handler, the
previous vCPU who starts searching earlier can mark IPI receivers in
user-mode as dy_eligible, the vCPU who starts searching a little later
can select it directly. However, after the above posting, the
PLE-caused vCPU should search the second full pass by himself.
Wanpeng
