Re: [PATCH] fix oops in updating thread cputime and task time

[Peter Zijlstra <peterz@infradead.org>]

On Sat, 2012-04-21 at 09:23 +0200, Fawzi Mohamed wrote:
> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> index 4603b9d..03a2d89 100644
> --- a/kernel/sched/core.c
> +++ b/kernel/sched/core.c
> @@ -2966,7 +2966,7 @@ void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
>                 u64 temp = (__force u64) rtime;
>  
>                 temp *= (__force u64) utime;
> -               do_div(temp, (__force u32) total);
> +               temp = div64_u64(temp, total);
>                 utime = (__force cputime_t) temp;
>         } else
>                 utime = rtime;
> @@ -2999,7 +2999,7 @@ void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
>                 u64 temp = (__force u64) rtime;
>  
>                 temp *= (__force u64) cputime.utime;
> -               do_div(temp, (__force u32) total);
> +               temp = div64_u64(temp, total);
>                 utime = (__force cputime_t) temp;
>         } else
>                 utime = rtime; 

I'm not entirely sure why it takes 19 days, suppose we have HZ=1000 and
your app never idles, it still takes 2^32/1000 seconds ~50 days to
overflow that u32.

Anyway, yes your patch avoids the /0 issue, but it leaves the other
problem with that code..

  rtime * utime / total

The multiplication can overflow the u64 at which point you're staring at
complete rubbish, this happens at about that same point.

So I figure we need to use the shiny new mult_frac() primitive.

32bit platforms are going to be staring at crap either way though, since
their entire time accounting (cputime_t) will start warping at that
point,.. not sure what if anything we should do about that though..
anybody?