Re: [PATCH 4/4] nohz: Fix iowait overcounting if iowait task migrates

[Denys Vlasenko <dvlasenk@redhat.com>]

On 04/24/2014 09:18 PM, Peter Zijlstra wrote:
> On Thu, Apr 24, 2014 at 08:45:58PM +0200, Denys Vlasenko wrote:
>> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
>> index 268a45e..ffea757 100644
>> --- a/kernel/sched/core.c
>> +++ b/kernel/sched/core.c
>> @@ -4218,7 +4218,14 @@ void __sched io_schedule(void)
>>  	current->in_iowait = 1;
>>  	schedule();
>>  	current->in_iowait = 0;
>> +#ifdef CONFIG_NO_HZ_COMMON
>> +	if (atomic_dec_and_test(&rq->nr_iowait)) {
>> +		if (raw_smp_processor_id() != cpu_of(rq))
>> +			tick_nohz_iowait_to_idle(cpu_of(rq));
>> +	}
>> +#else
>>  	atomic_dec(&rq->nr_iowait);
>> +#endif
>>  	delayacct_blkio_end();
>>  }
> 
> You're really refusing to collapse that stuff eh?

I'm sending two patches on top of my last patch set
which tidies up a few such aspects (another one
is where we fetch a percpu variable before knowing
that we'll need it, potentially wasting a few cycles).


>> +void tick_nohz_iowait_to_idle(int cpu)
>> +{
>> +	struct tick_sched *ts = tick_get_tick_sched(cpu);
>> +	ktime_t now = ktime_get();
>> +
>> +	write_seqcount_begin(&ts->idle_sleeptime_seq);
>> +	ts->iowait_exittime = now;
>> +	write_seqcount_end(&ts->idle_sleeptime_seq);
>> +}
>
>
> So what again was wrong with this one?
>
>   http://marc.info/?l=linux-kernel&m=139772917211023

That code has no provision to record when last iowait task
left the rq.

Therefore it can undercount iowait - it's very similar
to the problem I had before patch #4 in my patch series.

My patches 1-3 can overcount iowait because they consider
the entire idle period "iowait" if nr_iowait_cpu() != 0
at the *beginning*.

Hidetoshi's patches consider the entire idle period "iowait"
if nr_iowait_cpu() != 0 at the *end*.

He needs to code carefully so that this delayed decision
doesn't make reader functions return wrong results.

However, if nr_iowait_cpu() was 0 at the end it does not mean
that most of this time period it was also 0. It could have been
mostly !0 - and in this case iowait will be undercounted.

I personally thought that both over- or undercounting iowait
might be acceptable.

If not, then *some* form of recording and accounting
for exact moment when last iowait task left the rq is necessary.
That's what I did in patch #4.

-- 
vda