Re: [PATCH v2] KVM: x86: do not go through vcpu in __get_kvmclock_ns

From: Radim KrÄmÃÅ
Date: Wed Nov 16 2016 - 11:10:45 EST


2016-11-14 18:51+0100, Paolo Bonzini:
> Going through the first VCPU is wrong if you follow a KVM_SET_CLOCK with
> a KVM_GET_CLOCK immediately after, without letting the VCPU run and
> call kvm_guest_time_update.
>
> To fix this, compute the kvmclock value ourselves, using the master
> clock (tsc, nsec) pair as the base and the host CPU frequency as
> the scale.
>
> Reported-by: Marcelo Tosatti <mtosatti@xxxxxxxxxx>
> Signed-off-by: Paolo Bonzini <pbonzini@xxxxxxxxxx>
> ---
> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> @@ -1620,6 +1620,11 @@ static bool kvm_get_time_and_clockread(s64 *kernel_ns, cycle_t *cycle_now)
>
> return do_monotonic_boot(kernel_ns, cycle_now) == VCLOCK_TSC;
> }
> +#else
> +static inline bool kvm_get_time_and_clockread(s64 *kernel_ns, cycle_t *cycle_now)
> +{
> + return false;
> +}
> #endif

A left-over from v1. ;)

> @@ -1724,18 +1729,18 @@ static void kvm_gen_update_masterclock(struct kvm *kvm)
>
> static u64 __get_kvmclock_ns(struct kvm *kvm)
> {
> - struct kvm_vcpu *vcpu = kvm_get_vcpu(kvm, 0);
> struct kvm_arch *ka = &kvm->arch;
> - s64 ns;
> + struct pvclock_vcpu_time_info hv_clock;
>
> - if (vcpu->arch.hv_clock.flags & PVCLOCK_TSC_STABLE_BIT) {
> - u64 tsc = kvm_read_l1_tsc(vcpu, rdtsc());
> - ns = __pvclock_read_cycles(&vcpu->arch.hv_clock, tsc);
> - } else {
> - ns = ktime_get_boot_ns() + ka->kvmclock_offset;
> - }

If we access the "global" master clock, it would be better to prevent it
from changing under our hands with
spin_lock(&ka->pvclock_gtod_sync_lock).

> + if (!ka->use_master_clock)
> + return ktime_get_boot_ns() + ka->kvmclock_offset;
>
> - return ns;
> + hv_clock.tsc_timestamp = ka->master_cycle_now;
> + hv_clock.system_time = ka->master_kernel_ns + ka->kvmclock_offset;
> + kvm_get_time_scale(NSEC_PER_SEC, __this_cpu_read(cpu_tsc_khz) * 1000LL,
> + &hv_clock.tsc_shift,
> + &hv_clock.tsc_to_system_mul);

Doesn't this result in a minor drift with scaled clock, because the
guest can be combining two systems that approximate frequency?

1) tsc_shift and tsc_to_system_mul for kvmclock scaling
2) hardware TSC scaling ratio

If we are on a 7654321 kHz TSC and TSC-ratio scale to 1234567 kHz and
then tsc_shift+tsc_to_system_mul kvmclock-scale to 1000000 kHz, we
should be using multipliers of
0.161290204578564186163606151349022336533834941074459772460... and
0.810000591300431649315104000025920018921613812778083328000...,
to achieve that. Those multipliers cannot be precisely expressed in
what we have (shifts and 64/32 bit multipliers with intermediate values
only up to 128 bits), so performing the scaling will result in slightly
incorrect frequency.

The result of combining two operations that alter the freqency is quite
unlikely to cancel out and produce the same result as an operation that
uses a different shift+multiplier to scale in one step, so I think that
we aren't getting the same time as the guest with TSC-scaling is seeing.

(I'd be happier if we didn't ignore this drift when the whole endeavor
started just to get rid of a drift, but introducing a minor bug is still
improving the situation -- I'm ok with first two changes only.)

> + return __pvclock_read_cycles(&hv_clock, rdtsc());
> }
>
> u64 get_kvmclock_ns(struct kvm *kvm)
> --
> 1.8.3.1
>