Re: [PATCH] time/sched_clock: Allow architecture to override cyc_to_ns()

[John Stultz <john.stultz@linaro.org>]

On Mon, Nov 15, 2021 at 5:41 PM Huacai Chen <chenhuacai@gmail.com> wrote:
>
> Hi, John,
>
> On Tue, Nov 16, 2021 at 1:27 AM John Stultz <john.stultz@linaro.org> wrote:
> >
> > On Sat, Nov 13, 2021 at 11:47 PM Huacai Chen <chenhuacai@loongson.cn> wrote:
> > >
> > > The current cyc_to_ns() implementation is like this:
> > >
> > > static inline u64 notrace cyc_to_ns(u64 cyc, u32 mult, u32 shift)
> > > {
> > >         return (cyc * mult) >> shift;
> > > }
> > >
> > > But u64*u32 maybe overflow, so introduce ARCH_HAS_CYC_TO_NS to allow
> > > architecture to override it.
> > >
> >
> > If that's the case, it would seem too large a mult/shift pair had been selected.
> We use a 100MHz clock and the counter is 64bit, the mult is ~160M. But
> even if we use a smaller mult, cyc*mult, it can also overflow.

Well, yes, any simple multiplication could overflow. My point is that
the mult/shift pair is usually calculated for an expected interval
range via clocks_calc_mult_shift(), where the max interval for
sched_clock is set to an hour.  So any interval length under an hour
should not overflow in a simple multiplication.

What I'm trying to understand is what is the case you have where your
interval length is longer than an hour?
As that might break other assumptions going on in the code.

> > What sort of cycle range are you considering to be valid here? Can you
> > provide more rationale as to why this needs the ability to be
> > overridden?
> >
> > And what sort of arch-specific logic do you envision, rather than
> > having a common implementation to avoid the overflow?
> u64*u64 can be handled by hardware (store the high bits and low bits
> of result in two registers). So, if we use assembly, we can handle the
> overflow correctly. E.g., LoongArch (and MIPS) can override
> cyc_to_ns() like this:
>
> static inline u64 notrace cyc_to_ns(u64 cyc, u32 mult, u32 shift)
> {
>         u64 t1, t2, t3;
>         unsigned long long rv;
>
>         /* 64-bit arithmetic can overflow, so use 128-bit. */
>         __asm__ (
>                 "nor            %[t1], $r0, %[shift]    \n\t"
>                 "mulh.du        %[t2], %[cyc], %[mult]  \n\t"
>                 "mul.d          %[t3], %[cyc], %[mult]  \n\t"
>                 "slli.d         %[t2], %[t2], 1         \n\t"
>                 "srl.d          %[rv], %[t3], %[shift]  \n\t"
>                 "sll.d          %[t1], %[t2], %[t1]     \n\t"
>                 "or             %[rv], %[t1], %[rv]     \n\t"
>                 : [rv] "=&r" (rv), [t1] "=&r" (t1), [t2] "=&r" (t2),
> [t3] "=&r" (t3)
>                 : [cyc] "r" (cyc), [mult] "r" (mult), [shift] "r" (shift)
>                 : );
>         return rv;
> }

But then isn't the mul_u64_u32_shr() the right abstraction for such a
custom implementation?

Then potentially implement a generic cyc_to_ns() implementation that
uses that instead?

thanks
-john