Re: [PATCH v2 -next] sched/cputime: Fix mul_u64_u64_div_u64() precision for cputime

From: Peter Zijlstra
Date: Fri Jul 26 2024 - 09:08:46 EST


On Fri, Jul 26, 2024 at 03:04:01PM +0200, Oleg Nesterov wrote:
> On 07/26, Oleg Nesterov wrote:
> >
> > On 07/26, Zheng Zucheng wrote:
> > >
> > > before call mul_u64_u64_div_u64(),
> > > stime = 175136586720000, rtime = 135989749728000, utime = 1416780000.
> >
> > So stime + utime == 175138003500000
> >
> > > after call mul_u64_u64_div_u64(),
> > > stime = 135989949653530
> >
> > Hmm. On x86 mul_u64_u64_div_u64(175136586720000, 135989749728000, 175138003500000)
> > returns 135989749728000 == rtime, see below.
>
> Seriously, can you re-check your numbers? it would be nice to understand why
> x86_64 differs...

x86_64 has a custom mul_u64_u64_div_u64() implementation.

> > But perhaps it makes sense to improve the accuracy of mul_u64_u64_div_u64() ?
> > See the new() function in the code below.
>
> Just in case, the usage of ilog2 can be improved, but this is minor.

I meant to go look at this, it seems to loop less than your improved
version, but I'm chasing crashes atm. Perhaps it provides inspiration.

https://codebrowser.dev/llvm/compiler-rt/lib/builtins/udivmodti4.c.html#__udivmodti4