Re: [RFC PATCH v4 10/11] lib: vdso: Allow arches to override the ns shift operation

[Christophe Leroy]

Le 29/01/2020 Ã 08:14, Thomas Gleixner a ÃcritÂ:
> Andy Lutomirski <luto@xxxxxxxxxx> writes:
>
> > On Thu, Jan 16, 2020 at 11:57 AM Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:
> > > Andy Lutomirski <luto@xxxxxxxxxx> writes:
> > > > On Thu, Jan 16, 2020 at 9:58 AM Christophe Leroy
> > > >
> > > > Would mul_u64_u64_shr() be a good alternative?  Could we adjust it to
> > > > assume the shift is less than 32?  That function exists to benefit
> > > > 32-bit arches.
> > >
> > > We'd want mul_u64_u32_shr() for this. The rules for mult and shift are:
> >
> > That's what I meant to type...
>
> Just that it does not work. The math is:
>
>       ns = d->nsecs;   // That's the nsec value shifted left by d->shift
>
>       ns += ((cur - d->last) & d->mask) * mult;
>
>       ns >>= d->shift;
>
> So we cannot use mul_u64_u32_shr() because we need the addition there
> before shifting. And no, we can't drop the fractional part of
> d->nsecs. Been there, done that, got sporadic time going backwards
> problems as a reward. Need to look at that again as stuff has changed
> over time.
>
> On x86 we enforce that mask is 64bit, so the & operation is not there,
> but due to the nasties of TSC we have that conditional
>
>      if (cur > last)
>         return (cur - last) * mult;
>      return 0;
>
> Christophe, on PPC the decrementer/RTC clocksource masks are 64bit as
> well, so you can spare that & operation there too.

Yes, I did it already. It spares reading d->mast, that the main advantage.

Christophe