On Fri, 2022-07-29 at 21:14 +0000, Sean Christopherson wrote:
> On Thu, Jul 07, 2022, Simon Veith wrote:
> > The Time Stamp Counter (TSC) value register can be set to an absolute
> > value using the KVM_SET_MSRS ioctl, which calls kvm_synchronize_tsc()
> > internally.
> >
> > Since this is a per-vCPU register, and vCPUs are often managed by
> > separate threads, setting a uniform TSC value across all vCPUs is
> > challenging: After live migration, for example, the TSC value may need
> > to be adjusted to account for the migration downtime. Ideally, we would
> > want each vCPU to be adjusted by the same offset; but if we compute the
> > offset centrally, the TSC value may become out of date due to scheduling
> > delays by the time that each vCPU thread gets around to issuing
> > KVM_SET_MSRS.
> >
> > In preparation for the next patch, this change adds an optional, KVM
> > clock based time reference argument to kvm_synchronize_tsc(). This
> > argument, if present, is understood to mean "the TSC value being written
> > was valid at this corresponding KVM clock time point".
>
>
> Given that commit 828ca89628bf ("KVM: x86: Expose TSC offset controls to userspace")
> was merged less than a year ago, it would be helpful to explicitly call out why
> KVM_VCPU_TSC_CTRL doesn't work, and why that sub-ioctl can't be extended/modified to
> make it work.
>
> > kvm_synchronize_tsc() will then use this clock reference to adjust the
> > TSC value being written for any delays that have been incurred since the
> > provided TSC value was valid.
> >
> > Co-developed-by: David Woodhouse <dwmw@xxxxxxxxxxxx>
>
> Needs David's SOB.
For the record (and for the half-baked "do it like this, except in a
way that might stand a chance of compiling..." that I threw at Simon on
Slack and didn't necessarily even expect him to credit me for
authorship)
Signed-off-by: David Woodhouse <dwmw@xxxxxxxxxxxx>
Attachment:
smime.p7s
Description: S/MIME cryptographic signature
