Re: [PATCH] KVM: VMX: switch to hrtimer for TSC deadline timer when L2 guest is running

[Paolo Bonzini]

> On 07/06/16 16:01, Haozhong Zhang wrote:
> > On 07/06/16 09:46, Paolo Bonzini wrote:
> > > On 06/07/2016 08:05, Haozhong Zhang wrote:
> > > >> > Nested preemption timer is emulated by hrtimer, so it doesn't
> > > >> > influence vmcs02, why this is needed?
> > > > Nested (L2) preemption timer is not affected for sure, and this patch
> > > > is to fix another problem caused by using L1 preemption timer for L1
> > > > TSC deadline timer. When we use L1 VMX preemption timer for L1 TSC
> > > > deadline timer, we intend to use the pin-based exec config and the VMX
> > > > preemption timer value in vmcs01. However, when L2 guest is running,
> > > > vmcs02 is loaded as the current VMCS so that a different VMX
> > > > preemption timer config is used (i.e. VMX preemption timer is disabled
> > > > in prepare_vmcs02()). If we still use preemption timer for L1 TSC
> > > > deadline timer at this moment, then L1 TSC deadline timer will not be
> > > > able to be triggered when L2 guest is running.
> > > 
> > > The right fix then is to set the pin-based controls in prepare_vmcs02,
> > > based on vcpu->arch.hv_deadline_tsc.
> > 
> > Then KVM needs to distinguish L1 and L2 VMEXITs for preemption timer
> > (or does KVM already have this?). The current implementation which
> > uses a separate hrtimer for L2 preemption timer can easily distinguish
> > them.
> 
> And L1 hypervisor may also want to use preemption timer for L2
> guest. In this case, L0 KVM can not reuse preemption timer for both L1
> and L2.

The vmcs12's preemption timer is emulated by KVM through a separate hrtimer,
so KVM would only use the vmcs02 preemption timer for the L1 APIC timer.

Paolo