Re: [PATCH 2/4] sched: Account per task_group nr_iowait

[Peter Zijlstra]

On Mon, Nov 06, 2017 at 05:24:36PM +0100, Peter Zijlstra wrote:
> On Mon, Nov 06, 2017 at 07:12:58PM +0300, Kirill Tkhai wrote:
> 
> > >> +	atomic_inc(&tg->stat[rq->cpu].nr_iowait);
> > > 
> > > You're joking right, more atomic ops on the fast paths..
> > 
> > There should be a synchronization... It's modified under rq->lock everywhere, except try_to_wakeup().
> > Would it be better to use one more rq->lock at try_to_wakeup() instead of atomic?
> 
> No, of course not. We spend a lot of time getting of that rq->lock
                                                   ^ rid
> there.
> 
> The better option is to not care about iowait, since its a complete
> garbage number to begin with -- read that commit I pointed you to.
> 
> But if you do manage to convince me iowait is a sane thing to export
> (and its not); then you should not use atomics -- nor is there any need
> to. Since all you want to export is \Sum nr_iowait, you can inc/dec to
> pure cpu local variables and the sum will make it all work.
> 
> The extant iowait crap cannot do this because it thinks per-cpu IO-wait
> is a thing -- its not, its a random number at best, but its ABI so we
> can't fix :-(