Re: [PATCH v2 1/3] KVM: x86: implement KVM_{GET|SET}_TSC_STATE

[Maxim Levitsky]

On Tue, 2020-12-08 at 17:40 +0100, Thomas Gleixner wrote:
> On Tue, Dec 08 2020 at 13:13, Maxim Levitsky wrote:
> > On Mon, 2020-12-07 at 11:29 -0600, Oliver Upton wrote:
> > > How would a VMM maintain the phase relationship between guest TSCs
> > > using these ioctls?
> > 
> > By using the nanosecond timestamp. 
> >  
> > While I did made it optional in the V2 it was done for the sole sake of being 
> > able to set TSC on (re)boot to 0 from qemu, and for cases when qemu migrates 
> > from a VM where the feature is not enabled.
> > In this case the tsc is set to the given value exactly, just like you
> > can do today with KVM_SET_MSRS.
> > In all other cases the nanosecond timestamp will be given.
> >  
> > When the userspace uses the nanosecond timestamp, the phase relationship
> > would not only be maintained but be exact, even if TSC reads were not
> > synchronized and even if their restore on the target wasn't synchronized as well.
> >  
> > Here is an example:
> >  
> > Let's assume that TSC on source/target is synchronized, and that the guest TSC
> > is synchronized as well.
> > 
> > Let's call the guest TSC frequency F (guest TSC increments by F each second)
> >  
> > We do KVM_GET_TSC_STATE on vcpu0 and receive (t0,tsc0).
> > We do KVM_GET_TSC_STATE on vcpu1 after 1 second passed (exaggerated) 
> > and receive (t0 + 1s, tsc0 + F)
> 
> Why?
> 
> You freeeze the VM and store the realtime timestamp of doing that. At
> that point assuming a full sync host system the only interesting thing
> to store is the guest offset which is the same on all vCPUs and it is
> known already.
> 
> So on restore the only thing which needs to be adjusted is the guest
> wide offset.
> 
>      newoffset = oldoffset + (now - tfreeze)
> 
> Then set newoffset for all vCPUs. Anything else is complexity for no
> value and bound to fall apart in hard to debug ways.
> 
> The offset is still the same for all vCPUs whether you can restore them
> in the same nanosecond or whether you need 3 minutes for each one. It
> does not matter because when you restore vCPU1 3 minutes after vCPU0
> then TSC has advanced 3 minutes as well. It's still correct from the
> guest POV.
> 
> Even if you support TSCADJUST and let the guest write to it does not
> change the per guest offset at all. TSCADJUST is per [v]CPU and adds on
> top:
> 
>     tscvcpu = tsc_host + guest_offset + TSC_ADJUST
> 
> Scaling is just orthogonal and does not change any of this.

I agree with this, and I think that this is what we will end up doing.
Paulo, what do you think about this?

Best regards,
	Maxim Levitsky

> 
> Thanks,
> 
>         tglx
>