On Mon, 2024-04-08 at 17:43 -0700, Dongli Zhang wrote:
Hi Jack,
On 4/8/24 15:07, Jack Allister wrote:
This test proves that there is an inherent KVM/PV clock drift away from the guest TSC when KVM decides to update the PV time information structure due to a KVM_REQ_MASTERCLOCK_UPDATE. This drift is exascerbated when a guest is
Typo: exacerbated
using TSC scaling and running at a different frequency to the host TSC [1]. It also proves that KVM_[GS]ET_CLOCK_GUEST API is working to mitigate the drift from TSC to within ±1ns.
No, the KVM_[GS}ET_CLOCK_GUEST API is not about mitigating kernel bugs.
We *fix* kernel bugs, we don't make userspace work around them.
The KVM_[GS}ET_CLOCK_GUEST API allows userspace to perform accurate live update and live migration without disrupting the relationship between guest TSC and KVM clock.
Since a bombing run on KVM_REQ_MASTERCLOCK_UPDATE users is on my TODO list, it's worth noting the *reason* that switching to the obsolete MSR_KVM_SYSTEM_TIME forces ka->use_master_clock mode off.
It's not documented at all as far as I can tell, but in commit 54750f2cf042 (“KVM: x86: workaround SuSE's 2.6.16 pvclock vs masterclock issue“) in 2015, it was done to work around a guest bug where the guest *expected* the reference point to keep being updated and never be too far in the past.
Is configure_scaled_tsc() anecessary? Or how about to make it an option/arg? Then I will be able to test it on a VM/server without TSC scaling.
As discussed, TSC scaling shouldn't be needed. It *should* be part of the unit test if possible, because there is a class of bugs it'll trigger, but it should be optional.
In particular, the existing KVM_GET_CLOCK will return extra-wrong results if TSC scaling is in force. But that isn't being tested here yet because we haven't *fixed* it yet :)
For reference, here's my TODO list which Jack is working from...
• Add KVM unit test to validate that KVM clock does not change when provoked (including by simulated live update). It’s OK for the reference point at { tsc_timestamp, system_time } in the pvclock structure to change, but only such that it gives the same results for a given guest TSC — that is, if system_time changes, then tsc_timestamp must change by a delta which precisely corresponds in terms of the advertised guest TSC frequency. Perhaps allow a slop of 1ns for rounding, but no more.
• Audit and fix (i.e. remove) KVM_REQ_MASTERCLOCK_UPDATE usage, starting with kvm_xen_shared_info_init(). And work out whether it should be sent to all vCPUs, as some call sites do, or just one?
• Add KVM_VCPU_TSC_SCALE attribute to allow userspace to know the precise host→guest TSC scaling. (cf. https://lore.kernel.org/all/13f256ad95de186e3b6bcfcc1f88da5d0ad0cb71.camel@i...)
• Expose guest’s view of KVM clock to userspace via KVM_GET_CLOCK_GUEST ioctl. Perhaps also a memory-mapped version, as the gfn_to_pfn_cache allows writing to userspace HVAs. With this, userspace has fast and accurate way to calculate the KVM clock at any given moment in time. (Currently, userspace calls the KVM_GET_CLOCK ioctl which is slow and returns inaccurate results). Then userspace can base other things like PIT and HPET emulation on the KVM clock and simplify timekeeping over migration for those too.
• Add a KVM_SET_CLOCK_GUEST ioctl which consumes the pvclock information back again. This should not only set the kvmclock_offset field, but also set the reference point { master_cycle_now, master_kernel_ns } as follows: • Sample the kernel’s CLOCK_MONOTONIC_RAW to create a new master_kernel_ns and master_cycle_now. • Convert the new master_cycle_now to a guest TSC. • Calculate the intended KVM clock with that guest TSC from the provided pvclock information. • Calculate the current KVM clock with that guest TSC using the new master_cycle_now and master_kernel_ns and kvmclock_offset as usual. • Adjust kvmclock_offset to correct for the delta between current and intended values. • Raise KVM_REQ_CLOCK_UPDATE on all vCPUs.
• Fix the broken __get_kvmclock() function to scale via the guest’s TSC frequency as it should. There isn’t necessarily a vCPU to use for this, so it’s OK for this to work only when the frequency has been set of the whole VM rather than only for individual vCPUs. Likewise kvm_get_wall_clock_epoch() which has the same bug.
• Fix all other cases where KVM reads the time in two places separately and then treats them as simultaneous.
• Fix the discontinuities in KVM_REQ_MASTERCLOCK_UPDATE by allowing kvmclock_offset to vary while the VM is running in master clock mode. Perhaps every call to pvclock_update_vm_gtod_copy() which starts in master clock mode should follow the same process as the proposed KVM_SET_CLOCK_GUEST to adjust the kvmclock_offset value which corresponds with the new reference point. As long as we don’t break in the case where something weird (host hibernation, etc.) happened to the TSC, and we actually want to trust kvmclock_offset. Maybe we should have a periodic work queue which keeps kvmclock_offset in sync with the KVM clock while the VM is in master clock mode?
• Correct the KVM documentation for TSC migration to take TSC scaling into account. Something like...
(SOURCE) • Sample both TAI and the (source) host TSC at an arbitrary time we shall call Tsrc: • Use adjtimex() to obtain tai_offset. • Use KVM_GET_CLOCK to read UTC time and host TSC (ignoring the actual kvm clock). These represent time Tsrc. • Use adjtimex() to obtain tai_offset again, looping back to the beginning if it changes. • Convert the UTC time to TAI by adding the tai_offset.
• ∀ vCPU: • Read the scaling information with the KVM_CPU_TSC_SCALE attribute. • Read the offset with the KVM_CPU_TSC_OFFSET attribute. • Calculate this vCPU’s TSC at time Tsrc, from the host TSC value.
• Use KVM_GET_CLOCK_GUEST to read the KVM clock (on vCPU0).
(DESTINATION) • Sample both TAI and the (destination) host TSC at a time we shall call Tdst: • Use adjtimex() to obtain tai_offset. • Use KVM_GET_CLOCK to read UTC time and host TSC. • Use adjtimex() to obtain tai_offset again, looping back to the beginning if it changes. • Convert the UTC time to TAI by adding the tai_offset.
• Calculate the time (in the TAI clock) elapsed between Tsrc and Tdst. Call this ΔT.
• ∀ vCPU: • Calculate this vCPU’s intended TSC value at time Tdst: • Given this vCPU’s TSC frequency, calculate the number of TSC ticks correponding to ΔT. • Add this to the vCPU TSC value calculated on the source • Read the scaling information on the current host with the KVM_CPU_TSC_SCALE attribute • Calculate this vCPU’s scaled TSC value corresponding to the host TSC at time Tdst without taking offsetting into account. • Set KVM_CPU_TSC_OFFSET to the delta between that and the intended TSC value.
• Use KVM_SET_CLOCK_GUEST to set the KVM clock (on vCPU0).