Re: [PATCH] kvm: lapic: add module parameters for LAPIC_TIMER_ADVANCE_ADJUST_MAX/MIN

[Sean Christopherson <seanjc@google.com>]

On Wed, Mar 10, 2021, Haiwei Li wrote:
> On Wed, Mar 10, 2021 at 7:42 AM Sean Christopherson <seanjc@google.com> wrote:
> >
> > On Wed, Mar 03, 2021, Haiwei Li wrote:
> > > On 21/3/3 10:09, lihaiwei.kernel@gmail.com wrote:
> > > > From: Haiwei Li <lihaiwei@tencent.com>
> > > >
> > > > In my test environment, advance_expire_delta is frequently greater than
> > > > the fixed LAPIC_TIMER_ADVANCE_ADJUST_MAX. And this will hinder the
> > > > adjustment.
> > >
> > > Supplementary details:
> > >
> > > I have tried to backport timer related features to our production
> > > kernel.
> > >
> > > After completed, i found that advance_expire_delta is frequently greater
> > > than the fixed value. It's necessary to trun the fixed to dynamically
> > > values.
> >
> > Does this reproduce on an upstream kernel?  If so...
> >
> >   1. How much over the 10k cycle limit is the delta?
> >   2. Any idea what causes the large delta?  E.g. is there something that can
> >      and/or should be fixed elsewhere?
> >   3. Is it platform/CPU specific?
> 
> Hi, Sean
> 
> I have traced the flow on our production kernel and it frequently consumes more
> than 10K cycles from sched_out to sched_in.
> So two scenarios tested on Cascade lake Server(96 pcpu), v5.11 kernel.
> 
> 1. only cyclictest in guest(88 vcpu and bound with isolated pcpus, w/o mwait
> exposed, adaptive advance lapic timer is default -1). The ratio of occurrences:
> 
> greater_than_10k/total: 29/2060, 1.41%
> 
> 2. cyclictest in guest(88 vcpu and not bound, w/o mwait exposed, adaptive
> advance lapic timer is default -1) and stress in host(no isolate). The ratio of
> occurrences:
> 
> greater_than_10k/total: 122381/1017363, 12.03%

Hmm, I'm inclined to say this is working as intended.  If the vCPU isn't affined
and/or it's getting preempted, then large spikes are expected, and not adjusting
in reaction to those spikes is desirable.  E.g. adjusting by 20k cycles because
the timer happened to expire while a vCPU was preempted will cause KVM to busy
wait for quite a long time if the next timer runs without interference, and then
KVM will thrash the advancement.

And I don't really see the point in pushing the max adjustment beyond 10k.  The
max _advancement_ is 5000ns, which means that even with a blazing fast 5.0ghz
system, a max adjustment of 1250 (10k/ 8, the step divisor) should get KVM to
the 25000 cycle advancement limit relatively quickly.  Since KVM resets to the
initial 1000ns advancement when it would exceed the 5000ns max, I suspect that
raising the max adjustment much beyond 10k cycles would quickly push a vCPU to
the max, cause it to reset, and rinse and repeat.

Note, we definitely don't want to raise the 5000ns max, as waiting with IRQs
disabled for any longer than that will likely cause system instability.