Guest VMs can be provided with a para-virtualized clock source to perform timekeeping. A KVM guest can map in a PV clock via the MSR_KVM_SYSTEM_TIME/MSR_KVM_SYSTEM_TIME_NEW virtualized MSRs. Where as on a Xen guest this can be provided via the vcpu/shared info pages. These PV clocks both use a common structure which is mapped between host <-> guest to provide the PVTI (paravirtual time information) for the clock. This reference information is a guest TSC timestamp and a host system time at a singular point in time. Upon a live-update of a host or live-migration of an instance the PVTI may be recalculated by KVM. Using the existing KVM_[GS]ET_CLOCK functionality the relationship between the TSC and PV clock cannot be precisely saved and restored by userspace. This series adds in two patches, one to add in a new interface to allow a VMM/userspace to perform a correction of the PVTI structure. Then a second to verify the imprecision after a simulation of a live-update/migration and then to verify the correction is to within ±1ns. v1: https://lore.kernel.org/all/20240408220705.7637-1-jalliste@xxxxxxxxxx/ v2: - Moved new IOCTLs from vm to vcpu level. - Adds extra error checks as suggested by Dongli Zhang / David Woodhouse. - Adds on-demand calculation of PVTI if non currently present in vcpu. - Adds proper synchronization for PV clock during correction. - Added option to test without TSC scaling in sefltest. - Updated commit messages to better explain the situation (thanks David). Jack Allister (2): KVM: x86: Add KVM_[GS]ET_CLOCK_GUEST for accurate KVM clock migration KVM: selftests: Add KVM/PV clock selftest to prove timer correction Documentation/virt/kvm/api.rst | 37 ++++ arch/x86/kvm/x86.c | 124 +++++++++++ include/uapi/linux/kvm.h | 3 + tools/testing/selftests/kvm/Makefile | 1 + .../selftests/kvm/x86_64/pvclock_test.c | 192 ++++++++++++++++++ 5 files changed, 357 insertions(+) create mode 100644 tools/testing/selftests/kvm/x86_64/pvclock_test.c base-commit: 8cb4a9a82b21623dbb4b3051dd30d98356cf95bc -- 2.40.1
