Re: [tip:sched/core] sched: Lower chances of cputime scaling overflow

[Frederic Weisbecker <fweisbec@gmail.com>]

On Thu, Apr 11, 2013 at 08:22:03PM +0200, Frederic Weisbecker wrote:
> On Thu, Apr 11, 2013 at 08:07:50PM +0200, Peter Zijlstra wrote:
> > On Thu, 2013-04-11 at 08:38 -0700, Linus Torvalds wrote:
> > > On Thu, Apr 11, 2013 at 6:45 AM, Peter Zijlstra <peterz@infradead.org> wrote:
> > > > On Tue, 2013-03-26 at 15:01 +0100, Stanislaw Gruszka wrote:
> > > >> Thoughts?
> > > >
> > > > Would something like the below work?
> > > 
> > > Ugh, this is hard to think about, it's also fairly inefficient.
> > > 
> > > >  static cputime_t scale_stime(u64 stime, u64 rtime, u64 total)
> > > >  {
> > > > -       u64 rem, res, scaled;
> > > > +       int stime_fls = fls64(stime);
> > > > +       int total_fls = fls64(total);
> > > > +       int rtime_fls = fls64(rtime);
> > > 
> > > Doing "fls64()" unconditionally is quite expensive on some
> > > architectures, 
> > 
> > Oh, I (wrongly it appears) assumed that fls was something cheap :/
> > 
> > > and if I am not mistaken, the *common* case (by far) is
> > > that all these values fit in 32 bits, no?
> > 
> > It depends on if we use cputime_jiffies.h or cputime_nsec.h and I'm
> > completely lost as to which we default to atm. But we sure can reduce
> > to 32 bits in most cases without too much problems.
> 
> We default to the jiffies. The nsecs case is used only for full dynticks
> accounting and ia64 precise accounting.

Oh and in the latter case there is no scaling.