Re: [PATCH] Revert "KVM: x86: Unconditionally enable irqs in guest context"

[Sean Christopherson <seanjc@google.com>]

On Tue, Jan 12, 2021, Nitesh Narayan Lal wrote:
> 
> On 1/7/21 4:33 AM, Vitaly Kuznetsov wrote:
> > Sean Christopherson <seanjc@google.com> writes:
> >
> >> On Wed, Jan 06, 2021, Vitaly Kuznetsov wrote:
> >>> Looking back, I don't quite understand why we wanted to account ticks
> >>> between vmexit and exiting guest context as 'guest' in the first place;
> >>> to my understanging 'guest time' is time spent within VMX non-root
> >>> operation, the rest is KVM overhead (system).
> >> With tick-based accounting, if the tick IRQ is received after PF_VCPU is cleared
> >> then that tick will be accounted to the host/system.  The motivation for opening
> >> an IRQ window after VM-Exit is to handle the case where the guest is constantly
> >> exiting for a different reason _just_ before the tick arrives, e.g. if the guest
> >> has its tick configured such that the guest and host ticks get synchronized
> >> in a bad way.
> >>
> >> This is a non-issue when using CONFIG_VIRT_CPU_ACCOUNTING_GEN=y, at least with a
> >> stable TSC, as the accounting happens during guest_exit_irqoff() itself.
> >> Accounting might be less-than-stellar if TSC is unstable, but I don't think it
> >> would be as binary of a failure as tick-based accounting.
> >>
> > Oh, yea, I vaguely remember we had to deal with a very similar problem
> > but for userspace/kernel accounting. It was possible to observe e.g. a
> > userspace task going 100% kernel while in reality it was just perfectly
> > synchronized with the tick and doing a syscall just before it arrives
> > (or something like that, I may be misremembering the details).
> >
> > So depending on the frequency, it is probably possible to e.g observe
> > '100% host' with tick based accounting, the guest just has to
> > synchronize exiting to KVM in a way that the tick will always arrive
> > past guest_exit_irqoff().
> >
> > It seems to me this is a fundamental problem in case the frequency of
> > guest exits can match the frequency of the time accounting tick.
> >
> 
> Just to make sure that I am understanding things correctly.
> There are two issues:
> 1. The first issue is with the tick IRQs that arrive after PF_VCPU is
>    cleared as they are then accounted into the system context atleast on
>    the setup where CONFIG_VIRT_CPU_ACCOUNTING_GEN is not enabled. With the
>    patch "KVM: x86: Unconditionally enable irqs in guest context", we are
>    atleast taking care of the scenario where the guest context is exiting
>    constantly just before the arrival of the tick.

Yep.

> 2. The second issue that Sean mentioned was introduced because of moving
>    guest_exit_irqoff() closer to VM-exit. Due to this change, any ticks that
>    happen after IRQs are disabled are incorrectly accounted into the system
>    context. This is because we exit the guest context early without
>    ensuring if the required guest states to handle IRQs are restored.

Yep.

> So, the increase in the system time (reported by cpuacct.stats) that I was
> observing is not entirely correct after all.

It's correct, but iff CONFIG_VIRT_CPU_ACCOUNTING_GEN=y, as that doesn't rely on
ticks and so closer to VM-Enter is better.  The problem is that it completely
breaks CONFIG_VIRT_CPU_ACCOUNTING_GEN=n (#2 above) because KVM will never
service an IRQ, ticks included, with PF_VCPU set.

> Am I missing anything here?