> On Dec 7, 2020, at 8:38 AM, Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote: > > On Mon, Dec 07 2020 at 14:16, Maxim Levitsky wrote: >>> On Sun, 2020-12-06 at 17:19 +0100, Thomas Gleixner wrote: >>> From a timekeeping POV and the guests expectation of TSC this is >>> fundamentally wrong: >>> >>> tscguest = scaled(hosttsc) + offset >>> >>> The TSC has to be viewed systemwide and not per CPU. It's systemwide >>> used for timekeeping and for that to work it has to be synchronized. >>> >>> Why would this be different on virt? Just because it's virt or what? >>> >>> Migration is a guest wide thing and you're not migrating single vCPUs. >>> >>> This hackery just papers over he underlying design fail that KVM looks >>> at the TSC per vCPU which is the root cause and that needs to be fixed. >> >> I don't disagree with you. >> As far as I know the main reasons that kvm tracks TSC per guest are >> >> 1. cases when host tsc is not stable >> (hopefully rare now, and I don't mind making >> the new API just refuse to work when this is detected, and revert to old way >> of doing things). > > That's a trainwreck to begin with and I really would just not support it > for anything new which aims to be more precise and correct. TSC has > become pretty reliable over the years. > >> 2. (theoretical) ability of the guest to introduce per core tsc offfset >> by either using TSC_ADJUST (for which I got recently an idea to stop >> advertising this feature to the guest), or writing TSC directly which >> is allowed by Intel's PRM: > > For anything halfways modern the write to TSC is reflected in TSC_ADJUST > which means you get the precise offset. > > The general principle still applies from a system POV. > > TSC base (systemwide view) - The sane case > > TSC CPU = TSC base + TSC_ADJUST > > The guest TSC base is a per guest constant offset to the host TSC. > > TSC guest base = TSC host base + guest base offset > > If the guest want's this different per vCPU by writing to the MSR or to > TSC_ADJUST then you still can have a per vCPU offset in TSC_ADJUST which > is the offset to the TSC base of the guest. How about, if the guest wants to write TSC_ADJUST, it can turn off all paravirt features and keep both pieces?
