[Qemu-devel] Re: [PATCH] make tsc stable over migration and machine start

[Glauber Costa]

On Tue, 2011-02-01 at 21:26 +0100, Jan Kiszka wrote:
> On 2011-02-01 20:17, Glauber Costa wrote:
> > If the machine is stopped, we should not record two different tsc values
> > upon a save operation. The same problem happens with kvmclock.
> > 
> > But kvmclock is taking a different diretion, being now seen as a separate
> > device. Since this is unlikely to happen with the tsc, I am taking the
> > approach here of simply registering a handler for state change, and
> > using a per-CPUState variable that prevents double updates for the TSC.
> > 
> > Signed-off-by: Glauber Costa <address@hidden>
> > ---
> >  target-i386/cpu.h |    1 +
> >  target-i386/kvm.c |   19 ++++++++++++++++++-
> >  2 files changed, 19 insertions(+), 1 deletions(-)
> > 
> > diff --git a/target-i386/cpu.h b/target-i386/cpu.h
> > index 6d619e8..7f1c4f8 100644
> > --- a/target-i386/cpu.h
> > +++ b/target-i386/cpu.h
> > @@ -732,6 +732,7 @@ typedef struct CPUX86State {
> >      uint32_t sipi_vector;
> >      uint32_t cpuid_kvm_features;
> >      uint32_t cpuid_svm_features;
> > +    uint8_t  update_tsc;
> 
> bool please.
> 
> >      
> >      /* in order to simplify APIC support, we leave this pointer to the
> >         user */
> > diff --git a/target-i386/kvm.c b/target-i386/kvm.c
> > index ecb8405..c3925be 100644
> > --- a/target-i386/kvm.c
> > +++ b/target-i386/kvm.c
> > @@ -302,6 +302,16 @@ void kvm_inject_x86_mce(CPUState *cenv, int bank, 
> > uint64_t status,
> >  
> >  static int _kvm_arch_init_vcpu(CPUState *env);
> >  
> > +static void cpu_update_state(void *opaque, int running, int reason)
> > +{
> > +    CPUState *env = opaque;
> > +
> > +    if (!running) {
> > +        env->update_tsc = 1;
> > +    }
> > +}
> > +
> > +
> 
> Additional blank line.
> 
> >  int kvm_arch_init_vcpu(CPUState *env)
> >  {
> >      int r;
> > @@ -444,6 +454,8 @@ int kvm_arch_init_vcpu(CPUState *env)
> >      }
> >  #endif
> >  
> > +    qemu_add_vm_change_state_handler(cpu_update_state, env);
> > +
> >      return kvm_vcpu_ioctl(env, KVM_SET_CPUID2, &cpuid_data);
> >  }
> >  
> > @@ -1093,7 +1105,12 @@ static int kvm_get_msrs(CPUState *env)
> >     msrs[n++].index = MSR_STAR;
> >      if (kvm_has_msr_hsave_pa(env))
> >          msrs[n++].index = MSR_VM_HSAVE_PA;
> > -    msrs[n++].index = MSR_IA32_TSC;
> > +
> > +    if (env->update_tsc) {
> > +        msrs[n++].index = MSR_IA32_TSC;
> > +        env->update_tsc = 0;
> > +    }
> > +
> >  #ifdef TARGET_X86_64
> >      if (lm_capable_kernel) {
> >          msrs[n++].index = MSR_CSTAR;
> 
> Not quite the logic I'm using for kvmclock:

Ok. I have all the interest in keeping the same logic.
I will respin.

> cpu_update_state()
>       if running
>               tsc_valid = false;
> 
> kvm_get_msrs()
>       ...
>       if (!tsc_valid)
>               read_tsc
>               tsc_valid = !vm_running;
> 
> That ensure we always read the tsc while the VM is running, and not only
> after it was stopped (might otherwise be "surprising" when once
> visualizing the MSRs).
> 
> Jan
>