Re: idle time & iowait accounting

[Andrew Morton]

William Lee Irwin III wrote:
> 
> ...
> 
> +unsigned long nr_iowait(void)
> +{
> +       unsigned long i, sum = 0;
> +
> +       for (i = 0; i < NR_CPUS; ++i)
> +               sum += atomic_read(&cpu_rq(i)->nr_iowait);
> +
> +       return sum;
> +}
> +

We need to make a habit of checking cpu_online(i) in here.  I'd rather
this consume 4 cachelines than 32, thanks ;)

Also, if we decide to allocate each CPU's per-cpu memory separately
any for-all-CPUs loop which uses per-cpu data MUST make this check,
else it'll oops.

Not applicable in this case, and alas the cpu iterator helper macros
(for_each_online_cpu(), etc) are currently AWOL, but...

> +void io_schedule(void)
> +{
> +       struct runqueue *rq;
> +       preempt_disable();
> +       rq = this_rq();
> +       atomic_inc(&rq->nr_iowait);
> +       schedule();
> +       atomic_dec(&rq->nr_iowait);
> +       preempt_enable();
> +}

"scheduling while atomic".  You'll need to reacquire the runqueue pointer
on waking up.

> +
> +void io_schedule_timeout(long timeout)
> +{
> +       struct runqueue *rq;
> +       preempt_disable();
> +       rq = this_rq();
> +       atomic_inc(&rq->nr_iowait);
> +       schedule_timeout(timeout);
> +       atomic_dec(&rq->nr_iowait);
> +       preempt_enable();
> +}

And here too.

Apart from that, looks very good, thanks.