Re: [PATCH] Documentation: KVM: Describe guest TSC scaling in migration algorithm

[Oliver Upton <oupton@google.com>]

On Sat, Mar 19, 2022 at 08:52:56AM +0100, Paolo Bonzini wrote:
> On 3/18/22 19:39, Oliver Upton wrote:
> > Hi Paolo,
> > 
> > On Wed, Mar 16, 2022 at 08:47:59AM +0100, Paolo Bonzini wrote:
> > > On 3/16/22 05:53, Oliver Upton wrote:
> > > > The VMM has control of both the guest's TSC scale and offset. Extend the
> > > > described migration algorithm in the KVM_VCPU_TSC_OFFSET documentation
> > > > to cover TSC scaling.
> > > > 
> > > > Reported-by: David Woodhouse<dwmw@amazon.co.uk>
> > > > Signed-off-by: Oliver Upton<oupton@google.com>
> > > > ---
> > > > 
> > > > Applies to kvm/queue (references KVM_{GET,SET}_TSC_KHZ on a VM fd).
> > > 
> > > A few more things that have to be changed:
> > > 
> > > > 1. Invoke the KVM_GET_CLOCK ioctl to record the host TSC (tsc_src),
> > > >     kvmclock nanoseconds (guest_src), and host CLOCK_REALTIME nanoseconds
> > > >     (host_src).
> > > > 
> > > 
> > > One of two changes:
> > > 
> > > a) Add "Multiply tsc_src by guest_freq / src_freq to obtain scaled_tsc_src",
> > > add a new device attribute for the host TSC frequency.
> > > 
> > > b) Add "Multiply tsc_src by src_ratio to obtain scaled_tsc_src", add a new
> > > device attribute for the guest_frequency/host_frequency ratio.
> > > 
> > > A third would be scaling the host TSC frequency in KVM_GETCLOCK, but that's
> > > confusing IMO.
> > 
> > Agreed -- I think kvmclock should remain as is.
> > 
> > A fourth would be to expose the host's TSC frequency outside of KVM
> > since we're really in the business of guest features, not host ones :-)
> > We already have a patch that does this internally, and its visible in
> > some of our open source userspace libraries [1].
> 
> Yeah, it was a bit of a cop out on my part but adding it to sysfs would be
> nice.  Please tell me if any of you going to send a patch, or even just
> share it so that I can handle the upstream submission.

I'll have some time to look at it next week and send upstream. I was
somewhat panning if there were any other ideas here, but sysfs does seem
sensible :)

--
Thanks,
Oliver