Re: [PATCH v12 09/11] x86/tsc: prepare for early sched_clock

From: Pavel Tatashin
Date: Tue Jun 26 2018 - 15:48:36 EST


On Tue, Jun 26, 2018 at 2:42 PM Pavel Tatashin
<pasha.tatashin@xxxxxxxxxx> wrote:
>
> Hi Thomas,
>
> On Tue, Jun 26, 2018 at 11:44 AM Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:
> >
> > Pavel,
> >
> > first of all, sorry for my last outburst. I just was in a lousy mood after
> > staring into too much half baken stuff and failed to make myself stay away
> > from the computer.
>
> Thank you.
>
> >
> > On Sun, 24 Jun 2018, Thomas Gleixner wrote:
> > > On Sat, 23 Jun 2018, Pavel Tatashin wrote:
> > > And this early init sequence also needs to pull over the tsc adjust
> > > magic. So tsc_early_delay_calibrate() which should btw. be renamed to
> > > tsc_early_init() should have:
> > >
> > > {
> > > cpu_khz = x86_platform.calibrate_cpu();
> > > tsc_khz = x86_platform.calibrate_tsc();
> > >
> > > tsc_khz = tsc_khz ? : cpu_khz;
> > > if (!tsc_khz)
> > > return;
> > >
> > > /* Sanitize TSC ADJUST before cyc2ns gets initialized */
> > > tsc_store_and_check_tsc_adjust(true);
> > >
> > > calc_lpj(tsc_khz);
> > >
> > > tsc_sched_clock_init();
> > > }
> >
> > Peter made me look deeper into this and there are a few issues, which I
> > missed, depending on when some of the resources become available. So we
> > probably cannot hook all of this into tsc_early_delay_calibrate().
> >
> > I have an idea how to distangle it and we'll end up in a staged approach,
> > which looks like this:
> >
> > 1) Earliest one (not sure how early yet)
> >
> > Attempt to use MSR/CPUID. If not running on a hypervisor this can
> > try the quick PIT calibration, but nothing else.
> >
> > 2) Post init_hypervisor_platform()
> >
> > An attempt to use the hypervisor data can be made.
> >
> > 3) Post early_acpi_boot_init()
> >
> > This can do PIT/HPET based calibration
> >
> > 4) Post x86_dtb_init()
> >
> > PIT/PMTIMER based calibration
> >
> > Once tsc_khz is known, no further attempts of calibration are made. I'll
> > look into that later tonight.
>
> I think, there are no reasons to try staged attempts. It usually gets
> harder to maintain overtime. In my opinion it is best if do it in two
> tries, as right now, but just cleaner. The first attempt we get a
> crude result, using the lowest denominator to which current logic
> might fallback if something else is not available that early in boot:
> i.e cpu calibration loop in native_calibrate_cpu() but later get
> something better. Also, even if early clock does not work because we
> could not get tsc early, it is not a problem, we still will probably
> determine it later during tsc_init call.

Actually, nevermind, I looked through the code again, it seems that if
we get early tsc frequency we can keep it, but otherwise just try it
again at later time when in tsc_init(). So, no need for
cyc2ns_reinit_boot(). I still think no need for staged attempts, but
try in two different places: in tsc_init_early() -> works? use that
tsc frequency later, does not try again in tsc_init(), and use the new
one.

In tsc_init we can have something like this:
void __init tsc_init(void)
{
if (!boot_cpu_has(X86_FEATURE_TSC))
return;

/* See if we were not able to determine tsc frequency early,
but can now */
if (!tsc_khz && determine_cpu_tsc_frequncies()) {
/* Sanitize TSC ADJUST before cyc2ns gets initialized */
tsc_store_and_check_tsc_adjust(true);
cyc2ns_init_boot_cpu();
}
....

}

Pavel