Re: [patch, v2.6.22-rc6] sys_time() speedup

[Andrew Morton <akpm@linux-foundation.org>]

On Mon, 25 Jun 2007 23:09:46 +0200
Roman Zippel <zippel@linux-m68k.org> wrote:

> Hi,
> 
> On Monday 25 June 2007, Ingo Molnar wrote:
> 
> > the patch improves the sysbench OLTP macrobenchmark significantly:
> 
> Has that any real practical relevance?

Interesting question.  The patch adds a new test-n-branch to gettimeofday()
so if gettimeofday() is used much more frequently than time(), we lose.

> > @@ -373,6 +376,20 @@ void do_gettimeofday (struct timeval *tv
> >
> >  	tv->tv_sec = sec;
> >  	tv->tv_usec = usec;
> > +
> > +	/*
> > +	 * Make sure xtime.tv_sec [returned by sys_time()] always
> > +	 * follows the gettimeofday() result precisely. This
> > +	 * condition is extremely unlikely, it can hit at most
> > +	 * once per second:
> > +	 */
> > +	if (unlikely(xtime.tv_sec != tv->tv_sec)) {
> > +		unsigned long flags;
> > +
> > +		write_seqlock_irqsave(&xtime_lock);
> > +		update_wall_time();
> > +		write_seqlock_irqrestore(&xtime_lock);
> > +	}
> >  }
> >
> >  EXPORT_SYMBOL(do_gettimeofday);
> 
> Is this the do_gettimeofday() inside CONFIG_TIME_INTERPOLATION?

Yes.

> What did you test?
> There can be many ways to read the clock, do you want to put this hook 
> everywhere?

Yeah, it isn't immediately obvious (to this little black duck) why similar
fixups weren't needed in timekeeping.c.