Re: [PATCH 3/8] KVM: X86: Pass an additional 'L1' argument to kvm_scale_tsc()

["Stamatis, Ilias" <ilstam@amazon.com>]

On Mon, 2021-05-10 at 16:52 +0300, Maxim Levitsky wrote:
> On Thu, 2021-05-06 at 10:32 +0000, ilstam@mailbox.org wrote:
> > From: Ilias Stamatis <ilstam@amazon.com>
> > 
> > Sometimes kvm_scale_tsc() needs to use the current scaling ratio and
> > other times (like when reading the TSC from user space) it needs to use
> > L1's scaling ratio. Have the caller specify this by passing an
> > additional boolean argument.
> > 
> > Signed-off-by: Ilias Stamatis <ilstam@amazon.com>
> > ---
> >  arch/x86/include/asm/kvm_host.h |  2 +-
> >  arch/x86/kvm/x86.c              | 21 +++++++++++++--------
> >  2 files changed, 14 insertions(+), 9 deletions(-)
> > 
> > diff --git a/arch/x86/include/asm/kvm_host.h b/arch/x86/include/asm/kvm_host.h
> > index 132e820525fb..cdddbf0b1177 100644
> > --- a/arch/x86/include/asm/kvm_host.h
> > +++ b/arch/x86/include/asm/kvm_host.h
> > @@ -1779,7 +1779,7 @@ int kvm_pv_send_ipi(struct kvm *kvm, unsigned long ipi_bitmap_low,
> >  void kvm_define_user_return_msr(unsigned index, u32 msr);
> >  int kvm_set_user_return_msr(unsigned index, u64 val, u64 mask);
> > 
> > -u64 kvm_scale_tsc(struct kvm_vcpu *vcpu, u64 tsc);
> > +u64 kvm_scale_tsc(struct kvm_vcpu *vcpu, u64 tsc, bool l1);
> >  u64 kvm_read_l1_tsc(struct kvm_vcpu *vcpu, u64 host_tsc);
> > 
> >  unsigned long kvm_get_linear_rip(struct kvm_vcpu *vcpu);
> > diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> > index 7bc5155ac6fd..26a4c0f46f15 100644
> > --- a/arch/x86/kvm/x86.c
> > +++ b/arch/x86/kvm/x86.c
> > @@ -2241,10 +2241,14 @@ static inline u64 __scale_tsc(u64 ratio, u64 tsc)
> >       return mul_u64_u64_shr(tsc, ratio, kvm_tsc_scaling_ratio_frac_bits);
> >  }
> > 
> > -u64 kvm_scale_tsc(struct kvm_vcpu *vcpu, u64 tsc)
> > +/*
> > + * If l1 is true the TSC is scaled using L1's scaling ratio, otherwise
> > + * the current scaling ratio is used.
> > + */
> > +u64 kvm_scale_tsc(struct kvm_vcpu *vcpu, u64 tsc, bool l1)
> >  {
> >       u64 _tsc = tsc;
> > -     u64 ratio = vcpu->arch.tsc_scaling_ratio;
> > +     u64 ratio = l1 ? vcpu->arch.l1_tsc_scaling_ratio : vcpu->arch.tsc_scaling_ratio;
> > 
> >       if (ratio != kvm_default_tsc_scaling_ratio)
> >               _tsc = __scale_tsc(ratio, tsc);
> 
> I do wonder if it is better to have kvm_scale_tsc_l1 and kvm_scale_tsc instead for better
> readablility?
> 

That makes sense. Will do.

> 
> > @@ -2257,14 +2261,14 @@ static u64 kvm_compute_tsc_offset(struct kvm_vcpu *vcpu, u64 target_tsc)
> >  {
> >       u64 tsc;
> > 
> > -     tsc = kvm_scale_tsc(vcpu, rdtsc());
> > +     tsc = kvm_scale_tsc(vcpu, rdtsc(), true);
> 
> Here we have a somewhat dangerous assumption that this function
> will always be used with L1 tsc values.
> 
> The kvm_compute_tsc_offset should at least be renamed to
> kvm_compute_tsc_offset_l1 or something like that.
> 
> Currently the assumption holds though:
> 
> We call the kvm_compute_tsc_offset in:
> 
> -> kvm_synchronize_tsc which is nowadays thankfully only called
> on TSC writes from qemu, which are assumed to be L1 values.
> 
> (this is pending a rework of the whole thing which I started
> some time ago but haven't had a chance to finish it yet)
> 
> -> Guest write of MSR_IA32_TSC. The value written is in L1 units,
> since TSC offset/scaling only covers RDTSC and RDMSR of the IA32_TSC,
> while WRMSR should be intercepted by L1 and emulated.
> If it is not emulated, then L2 would just write L1 value.
> 
> -> in kvm_arch_vcpu_load, when TSC is unstable, we always try to resume
> the guest from the same TSC value as it had seen last time,
> and then catchup.

Yes. I wasn't sure about kvm_compute_tsc_offset but my understanding was
that all of its callers wanted the L1 value scaled.

Renaming it to kvm_scale_tsc_l1 sounds like a great idea.

> Also host TSC values are used, and after reading this function,
> I recommend to rename the vcpu->arch.last_guest_tsc
> to vcpu->arch.last_guest_tsc_l1 to document this.

OK

> > 
> >       return target_tsc - tsc;
> >  }
> > 
> >  u64 kvm_read_l1_tsc(struct kvm_vcpu *vcpu, u64 host_tsc)
> >  {
> > -     return vcpu->arch.l1_tsc_offset + kvm_scale_tsc(vcpu, host_tsc);
> > +     return vcpu->arch.l1_tsc_offset + kvm_scale_tsc(vcpu, host_tsc, true);
> 
> OK
> >  }
> >  EXPORT_SYMBOL_GPL(kvm_read_l1_tsc);
> > 
> > @@ -2395,9 +2399,9 @@ static inline void adjust_tsc_offset_guest(struct kvm_vcpu *vcpu,
> > 
> >  static inline void adjust_tsc_offset_host(struct kvm_vcpu *vcpu, s64 adjustment)
> >  {
> > -     if (vcpu->arch.tsc_scaling_ratio != kvm_default_tsc_scaling_ratio)
> > +     if (vcpu->arch.l1_tsc_scaling_ratio != kvm_default_tsc_scaling_ratio)
> >               WARN_ON(adjustment < 0);
> 
> This should belong to patch 2 IMHO.
> 

Right, I will move it.

> > -     adjustment = kvm_scale_tsc(vcpu, (u64) adjustment);
> > +     adjustment = kvm_scale_tsc(vcpu, (u64) adjustment, true);
> 
> OK
> >       adjust_tsc_offset_guest(vcpu, adjustment);
> >  }
> > 
> > @@ -2780,7 +2784,7 @@ static int kvm_guest_time_update(struct kvm_vcpu *v)
> >       /* With all the info we got, fill in the values */
> > 
> >       if (kvm_has_tsc_control)
> > -             tgt_tsc_khz = kvm_scale_tsc(v, tgt_tsc_khz);
> > +             tgt_tsc_khz = kvm_scale_tsc(v, tgt_tsc_khz, true);
> 
> OK (kvmclock is for L1 only, L1 hypervisor is free to expose its own kvmclock to L2)
> > 
> >       if (unlikely(vcpu->hw_tsc_khz != tgt_tsc_khz)) {
> >               kvm_get_time_scale(NSEC_PER_SEC, tgt_tsc_khz * 1000LL,
> > @@ -3474,7 +3478,8 @@ int kvm_get_msr_common(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
> >               u64 tsc_offset = msr_info->host_initiated ? vcpu->arch.l1_tsc_offset :
> >                                                           vcpu->arch.tsc_offset;
> > 
> > -             msr_info->data = kvm_scale_tsc(vcpu, rdtsc()) + tsc_offset;
> > +             msr_info->data = kvm_scale_tsc(vcpu, rdtsc(),
> > +                                            msr_info->host_initiated) + tsc_offset;
> 
> Since we now do two things that depend on msr_info->host_initiated, I
> think I would prefer to convert this back to regular 'if'.
> I don't have a strong opinion on this though.
> 

Agreed.

Thanks!

Ilias

> 
> >               break;
> >       }
> >       case MSR_MTRRcap:
> 
> 
> Best regards,
>         Maxim Levitsky
> 
>