Re: [PATCH v2] sched/cputime: make scale_stime() more precise

[Peter Zijlstra]

On Tue, May 19, 2020 at 11:33:34AM -0700, Linus Torvalds wrote:
> On Tue, May 19, 2020 at 10:25 AM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> >
> > --- a/include/linux/math64.h
> > +++ b/include/linux/math64.h
> > @@ -263,6 +263,47 @@ static inline u64 mul_u64_u32_div(u64 a, u32 mul, u32 divisor)
> >  }
> >  #endif /* mul_u64_u32_div */
> >
> > +#ifndef mul_u64_u64_div_u64
> > +static inline u64 mul_u64_u64_div_u64(u64 a, u64 b, u64 c)
> 
> Do we really want to inline this? Particularly if we expect this to be
> the "architecture doesn't have a better version" case?
> 
> It's not like we'd expect people to call it with constant values that
> could be optimized by inlining, do we? If any of the values are
> actually constants and it's performance-critical, the code is likely
> better off using some special helper rather than this anyway.
> 
> So I'd much rather see it as a weak function defined in
> lib/math/div64.c, and then architectures don't even need to override
> it at all. Just let them define their own (inline or not) function,
> and we have this as a weak fallback.

I completely forgot we had a .c file to go with all this. Yes, I'll put
it in there.