Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Date:   Tue, 12 Jan 2021 14:04:38 -0800
From:   Sean Christopherson <seanjc@google.com>
To:     Nitesh Narayan Lal <nitesh@redhat.com>
Cc:     Vitaly Kuznetsov <vkuznets@redhat.com>,
        linux-kernel@vger.kernel.org, kvm@vger.kernel.org,
        w90p710@gmail.com, pbonzini@redhat.com,
        Thomas Gleixner <tglx@linutronix.de>
Subject: Re: [PATCH] Revert "KVM: x86: Unconditionally enable irqs in guest
 context"
Message-ID: <X/4c9skP3rOqsWHW@google.com>
References: <20210105192844.296277-1-nitesh@redhat.com>
 <874kjuidgp.fsf@vitty.brq.redhat.com>
 <X/XvWG18aBWocvvf@google.com>
 <87ble1gkgx.fsf@vitty.brq.redhat.com>
 <fb41e24f-a5e3-8319-d25b-e0fe6b902a2b@redhat.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=iso-8859-1
Content-Disposition: inline
Content-Transfer-Encoding: 8bit
In-Reply-To: <fb41e24f-a5e3-8319-d25b-e0fe6b902a2b@redhat.com>
Precedence: bulk

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?