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

[Hidetoshi Seto <seto.hidetoshi@jp.fujitsu.com>]

(2014/04/24 4:00), Denys Vlasenko wrote:
> Before this change, if last IO-blocked task wakes up
> on a different CPU, the original CPU may stay idle for much longer,
> and the entire time it stays idle is accounted as iowait time.
> 
> This change adds struct tick_sched::iowait_exittime member.
> On entry to idle, it is set to KTIME_MAX.
> Last IO-blocked task, if migrated, sets it to current time.
> Note that this can happen only once per each idle period:
> new iowaiting tasks can't magically appear on idle CPU's rq.
> 
> If iowait_exittime is set, then (iowait_exittime - idle_entrytime)
> gets accounted as iowait, and the remaining (now - iowait_exittime)
> as "true" idle.
> 
> Run-tested: /proc/stats no longer go backwards.
> 
> Signed-off-by: Denys Vlasenko <dvlasenk@redhat.com>
> Cc: Frederic Weisbecker <fweisbec@gmail.com>
> Cc: Hidetoshi Seto <seto.hidetoshi@jp.fujitsu.com>
> Cc: Fernando Luis Vazquez Cao <fernando_b1@lab.ntt.co.jp>
> Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
> Cc: Thomas Gleixner <tglx@linutronix.de>
> Cc: Ingo Molnar <mingo@kernel.org>
> Cc: Peter Zijlstra <peterz@infradead.org>
> Cc: Andrew Morton <akpm@linux-foundation.org>
> Cc: Arjan van de Ven <arjan@linux.intel.com>
> Cc: Oleg Nesterov <oleg@redhat.com>

At first I'd like to say thank you very much for continuing
work on this problem.

My comments are inlined.

> ---
>  include/linux/tick.h     |  3 ++
>  kernel/sched/core.c      | 20 ++++++++++++++
>  kernel/time/tick-sched.c | 72 ++++++++++++++++++++++++++++++++++++++++++++----
>  3 files changed, 89 insertions(+), 6 deletions(-)
> 
> diff --git a/include/linux/tick.h b/include/linux/tick.h
> index b84773c..2a27883 100644
> --- a/include/linux/tick.h
> +++ b/include/linux/tick.h
> @@ -67,6 +67,7 @@ struct tick_sched {
>  	ktime_t				idle_exittime;
>  	ktime_t				idle_sleeptime;
>  	ktime_t				iowait_sleeptime;
> +	ktime_t				iowait_exittime;
>  	ktime_t				sleep_length;
>  	unsigned long			last_jiffies;
>  	unsigned long			next_jiffies;
> @@ -139,6 +140,7 @@ extern void tick_nohz_irq_exit(void);
>  extern ktime_t tick_nohz_get_sleep_length(void);
>  extern u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time);
>  extern u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time);
> +extern void tick_nohz_iowait_to_idle(int cpu);
>  
>  # else /* !CONFIG_NO_HZ_COMMON */
>  static inline int tick_nohz_tick_stopped(void)
> @@ -157,6 +159,7 @@ static inline ktime_t tick_nohz_get_sleep_length(void)
>  }
>  static inline u64 get_cpu_idle_time_us(int cpu, u64 *unused) { return -1; }
>  static inline u64 get_cpu_iowait_time_us(int cpu, u64 *unused) { return -1; }
> +static inline void tick_nohz_iowait_to_idle(int cpu) { }
>  # endif /* !CONFIG_NO_HZ_COMMON */
>  
>  #ifdef CONFIG_NO_HZ_FULL
> 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;
>  }

As I already mentioned in previous discussion with Peter, I have
concern here that this change might have impact on performance.
Especially in case if system is a kind of io-busy box, originally
there may be no iowait time (and possibly also no idle time).
For such case this change adds extra execution cost to manage
value of iowait_exittime which might not used.

Did you have any benchmark on this?

(And if you have extra time, it would be helpful if you could
 compare performance impacts of your set and mine:
  [PATCH v4 0/2] nohz: fix idle accounting in NO_HZ kernels
  https://lkml.org/lkml/2014/4/17/120
)

And if we successfully found a way to get the iowait_exittime
within reasonable negligible cheap cost, then why we don't use
it for NOHZ=n kernels too? 

> diff --git a/kernel/time/tick-sched.c b/kernel/time/tick-sched.c
> index a99770e..679a577 100644
> --- a/kernel/time/tick-sched.c
> +++ b/kernel/time/tick-sched.c
> @@ -403,17 +403,48 @@ static void tick_nohz_update_jiffies(ktime_t now)
>  	touch_softlockup_watchdog();
>  }
>  
> +/*
> + * Race prevention for idle/iowait counters:
> + *
> + * On entry and exit to idle/iowait states,
> + * ts->idle_entrytime, ts->{idle,iowait}_sleeptime and ts->idle_active
> + * fields are updated in this order, serialized with smp_wmb().
> + * On exit, ts->idle_entrytime is zeroed.
> + *
> + * The readers - get_cpu_{idle,iowait}_time_us() - fetch these fields
> + * in reversed order, serialized with smp_rmb().
> + * If ts->idle_active == 0, reading of ts->idle_entrytime (and one smp_rmb())
> + * are not necessary. If ts->idle_active != 0, ts->idle_entrytime
> + * (which is needed anyway for calculations)
> + * needs to be checked for zero, and reads need to be done anew if it is zeroed.
> + *
> + * This implements a seqlock-like race detection without needing additional
> + * data fields and read-modify-write ops on them.
> + */
>  static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now)
>  {
> -	ktime_t delta;
> +	ktime_t delta, entry;
>  
>  	/* Updates the per cpu time idle statistics counters */
> -	delta = ktime_sub(now, ts->idle_entrytime);
> +	entry = ts->idle_entrytime;
> +	delta = ktime_sub(now, entry);
>  	ts->idle_entrytime.tv64 = 0;
>  	smp_wmb();
>  
> -	if (ts->idle_active == 2)
> +	if (ts->idle_active == 2) {
> +		ktime_t end = ts->iowait_exittime;
> +
> +		if (end.tv64 != KTIME_MAX) {
> +			/*
> +			 * Last iowaiting task on our rq was woken up on other CPU
> +			 * sometime in the past, it updated ts->iowait_exittime.
> +			 */
> +			delta = ktime_sub(now, end);
> +			ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
> +			delta = ktime_sub(end, entry);
> +		}
>  		ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
> +	}
>  	else
>  		ts->idle_sleeptime = ktime_add(ts->idle_sleeptime, delta);
>  
> @@ -430,10 +461,18 @@ static ktime_t tick_nohz_start_idle(struct tick_sched *ts)
>  	ts->idle_entrytime = now;
>  	smp_wmb();
>  	ts->idle_active = nr_iowait_cpu(smp_processor_id()) ? 2 : 1;
> +	ts->iowait_exittime.tv64 = KTIME_MAX;
>  	sched_clock_idle_sleep_event();
>  	return now;
>  }
>  
> +void tick_nohz_iowait_to_idle(int cpu)
> +{
> +	struct tick_sched *ts = tick_get_tick_sched(cpu);
> +
> +	ts->iowait_exittime = ktime_get();
> +}
> +
>  /**
>   * get_cpu_idle_time_us - get the total idle time of a cpu
>   * @cpu: CPU number to query
> @@ -467,7 +506,7 @@ u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
>  	active = ACCESS_ONCE(ts->idle_active);
>  	smp_rmb();
>  	count = ACCESS_ONCE(ts->idle_sleeptime);
> -	if (active == 1) {
> +	if (active /* either 1 or 2 */) {
>  		ktime_t delta, start;
>  
>  		smp_rmb();
> @@ -479,6 +518,18 @@ u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
>  			 */
>  			goto again;
>  		}
> +		if (active == 2) {
> +			/* This idle period started as "iowait idle" */
> +			ktime_t iowait_exit = ACCESS_ONCE(ts->iowait_exittime);
> +
> +			if (iowait_exit.tv64 == KTIME_MAX)
> +				goto skip; /* and it still is */
> +			/*
> +			 * This CPU used to be "iowait idle", but iowait task
> +			 * has migrated. The rest of idle time is "true idle":
> +			 */
> +			start = iowait_exit;
> +		}
>  		delta = ktime_sub(now, start);
>  		count = ktime_add(count, delta);
>  	} else {
> @@ -488,6 +539,7 @@ u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time)
>  		 * we fetched idle_active, so count must be valid.
>  		 */
>  	}
> + skip:
>  
>  	return ktime_to_us(count);
>  }
> @@ -527,7 +579,7 @@ u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
>  	smp_rmb();
>  	count = ACCESS_ONCE(ts->iowait_sleeptime);
>  	if (active == 2) {
> -		ktime_t delta, start;
> +		ktime_t delta, start, end;
>  
>  		smp_rmb();
>  		start = ACCESS_ONCE(ts->idle_entrytime);
> @@ -538,7 +590,15 @@ u64 get_cpu_iowait_time_us(int cpu, u64 *last_update_time)
>  			 */
>  			goto again;
>  		}
> -		delta = ktime_sub(now, start);
> +		/*
> +		 * If last iowaiting task on our rq was woken up on other CPU
> +		 * sometime in the past, it updated ts->iowait_exittime.
> +		 * Otherwise, ts->iowait_exittime == KTIME_MAX.
> +		 */
> +		end = ACCESS_ONCE(ts->iowait_exittime);
> +		if (end.tv64 == KTIME_MAX)
> +			end = now;
> +		delta = ktime_sub(end, start);
>  		count = ktime_add(count, delta);
>  	} else {
>  		/*
> 

As Frederic already pointed, seqcount must be better choice.


Thanks,
H.Seto