Re: [PATCH] Fix almost infinite loop in APIC

[Sheng Yang <sheng@linux.intel.com>]

On Wednesday 21 January 2009 02:51:39 Marcelo Tosatti wrote:
> Hi Sheng,
>
> On Tue, Jan 20, 2009 at 09:43:15PM +0800, Sheng Yang wrote:
> > Marcelo, I realize some interesting things.
> >
> > In fact, on my machine, when I measured the delta of now() and
> > last_update at kvm_apic_timer_intr_post(), the delta was bigger and
> > bigger...(and now() is always bigger for me, so still no clue for why
> > data can be like above, last_update always ahead of when for about one
> > period...).
> >
> > Then I found something not good in original design - it ignored the
> > interval between time fire and injection, so we got:
> >
> > last_update = now() + n * period;
> >
> > And the time we update last_update:
> >
> > time = now() + n * period + n * interval.
> >
> > So last_update time is more and more inaccurate... Though it was revised
> > by tmcct function, it's still not a good way to go.
> >
> > Then I understand your purpose more.
> >
> > +       if (unlikely(atomic_read(&apic->timer.pending) > 0)) {
> > +               remaining = apic->timer.injection_delay;
> > +               if (ktime_to_ns(remaining) > apic->timer.period)
> > +                       remaining = ns_to_ktime(apic->timer.period);
> > +        } else
> > +               remaining = hrtimer_expires_remaining(&apic->timer.dev);
> >
> > And about your patch, how about take interval between intr_post() and
> > read tmcct in to account as well?
>
> You mean to sum up interval between intr_post() and read tmcct into
> injection_delay?
>
> I don't get it. Can you be more descriptive please?

Mind continue to change...

A: time_fire
B: intr_post
C: read TMCCT

I think the sequence can be AABC or AACB(I just consider AABC last night...). 
We have determined to return A->C as TMCCT, and the interval between different 
A is period which can be ignored. So how about simply return 
hrtimer_expires_remaining() for all tmcct reading including pending ones?

(And sorry for I still can't get why you return injection delay for TMCCT...)

-- 
regards
Yang, Sheng