Re: [PATCH 3/3] nohz: Fix iowait overcounting if iowait task migrates

[Denys Vlasenko <dvlasenk@redhat.com>]

On 04/24/2014 09:07 AM, Peter Zijlstra wrote:
> On Wed, Apr 23, 2014 at 09:00:36PM +0200, Denys Vlasenko wrote:
>> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
>> index 9cae286..08dd220 100644
>> --- a/kernel/sched/core.c
>> +++ b/kernel/sched/core.c
>> @@ -4255,6 +4255,9 @@ EXPORT_SYMBOL_GPL(yield_to);
>>   */
>>  void __sched io_schedule(void)
>>  {
>> +#ifdef CONFIG_NO_HZ_COMMON
>> +	int cpu_on_entry = smp_processor_id();
>> +#endif
>>  	struct rq *rq = raw_rq();
>>  
>>  	delayacct_blkio_start();
>> @@ -4263,13 +4266,23 @@ 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 (smp_processor_id() != cpu_on_entry)
>> +			tick_nohz_iowait_to_idle(cpu_on_entry);
>> +	}
>> +#else
>>  	atomic_dec(&rq->nr_iowait);
>> +#endif
>>  	delayacct_blkio_end();
>>  }
>>  EXPORT_SYMBOL(io_schedule);
>>  
>>  long __sched io_schedule_timeout(long timeout)
>>  {
>> +#ifdef CONFIG_NO_HZ_COMMON
>> +	int cpu_on_entry = smp_processor_id();
>> +#endif
>>  	struct rq *rq = raw_rq();
>>  	long ret;
>>  
>> @@ -4279,7 +4292,14 @@ long __sched io_schedule_timeout(long timeout)
>>  	current->in_iowait = 1;
>>  	ret = schedule_timeout(timeout);
>>  	current->in_iowait = 0;
>> +#ifdef CONFIG_NO_HZ_COMMON
>> +	if (atomic_dec_and_test(&rq->nr_iowait)) {
>> +		if (smp_processor_id() != cpu_on_entry)
>> +			tick_nohz_iowait_to_idle(cpu_on_entry);
>> +	}
>> +#else
>>  	atomic_dec(&rq->nr_iowait);
>> +#endif
>>  	delayacct_blkio_end();
>>  	return ret;
>>  }
> 
> Why do you insist on writing the same (buggy, see later) code twice?

I don't insist on such a thing. I am modifying existing code
which already has these nearly-identical functions.

I agree with you that they beg for common code being moved into
a shared helper function.


> Its buggy because the smp_processor_id() is used in preemptible context,
> its further buggy because the raw_rq() does it again and could get a rq
> on a different cpu.
> 
> What you want is something like:
> 
> static inline void io_wait_start(struct rq *rq)
> {
> 	atomic_inc(&rq->nr_iowait);
> 	current->in_iowait = 1;
> }
> 
> static inline void io_wait_end(struct rq *rq)
> {
> 	current->in_iowait = 0;
> #ifdef CONFIG_NO_HZ_COMMON
> 	if (atomic_dec_and_test(&rq->nr_iowait) && 
> 		cpu_of(rq) != raw_smp_processor_id()) {
> 		tick_nohz_iowait_end(cpu_of(rq));
> 	}
> #else
> 	atomic_dec(&rq->nr_iowait);
> #endif
> }

Thanks, will use this in the next patchset.