Re: [PATCH clocksource 2/2] clocksource: Exponential backoff for load-induced bogus watchdog reads

[Waiman Long]

On 11/2/22 14:40, Paul E. McKenney wrote:
> The clocksource watchdog will reject measurements that are excessively
> delayed, that is, by more than 1.5 seconds beyond the intended 0.5-second
> watchdog interval.  On an extremely busy system, this can result in a
> console message being printed every two seconds.  This is excessively
> noisy for a non-error condition.
>
> Therefore, apply exponential backoff to these messages.  This exponential
> backoff is capped at 1024 times the watchdog interval, which comes to
> not quite one message per ten minutes.
>
> Please note that the bogus watchdog reads that occur when the watchdog
> interval is less than 0.125 seconds are still printed unconditionally
> because these likely correspond to a serious error condition in the
> timer code or hardware.
>
> [ paulmck: Apply Feng Tang feedback. ]
>
> Reported-by: Waiman Long <longman@xxxxxxxxxx>
> Reported-by: Feng Tang <feng.tang@xxxxxxxxx>
> Signed-off-by: Paul E. McKenney <paulmck@xxxxxxxxxx>
> Cc: John Stultz <jstultz@xxxxxxxxxx>
> Cc: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
> Cc: Stephen Boyd <sboyd@xxxxxxxxxx>
> Cc: Feng Tang <feng.tang@xxxxxxxxx>
> Cc: Waiman Long <longman@xxxxxxxxxx>
> ---
>   include/linux/clocksource.h |  4 ++++
>   kernel/time/clocksource.c   | 31 +++++++++++++++++++++++++------
>   2 files changed, 29 insertions(+), 6 deletions(-)
>
> diff --git a/include/linux/clocksource.h b/include/linux/clocksource.h
> index 1d42d4b173271..daac05aedf56a 100644
> --- a/include/linux/clocksource.h
> +++ b/include/linux/clocksource.h
> @@ -125,6 +125,10 @@ struct clocksource {
>   	struct list_head	wd_list;
>   	u64			cs_last;
>   	u64			wd_last;
> +	u64			wd_last_bogus;
> +	int			wd_bogus_shift;
> +	unsigned long		wd_bogus_count;
> +	unsigned long		wd_bogus_count_last;
>   #endif
>   	struct module		*owner;
>   };
> diff --git a/kernel/time/clocksource.c b/kernel/time/clocksource.c
> index 3f5317faf891f..de8047b6720f5 100644
> --- a/kernel/time/clocksource.c
> +++ b/kernel/time/clocksource.c
> @@ -442,14 +442,33 @@ static void clocksource_watchdog(struct timer_list *unused)
>   
>   		/* Check for bogus measurements. */
>   		wdi = jiffies_to_nsecs(WATCHDOG_INTERVAL);
> -		if (wd_nsec < (wdi >> 2)) {
> -			/* This usually indicates broken timer code or hardware. */
> -			pr_warn("timekeeping watchdog on CPU%d: Watchdog clocksource '%s' advanced only %lld ns during %d-jiffy time interval, skipping watchdog check.\n", smp_processor_id(), watchdog->name, wd_nsec, WATCHDOG_INTERVAL);
> +		if (wd_nsec > (wdi << 2) || cs_nsec > (wdi << 2)) {
> +			bool needwarn = false;
> +			u64 wd_lb;
> +
> +			cs->wd_bogus_count++;
> +			if (!cs->wd_bogus_shift) {
> +				needwarn = true;
> +			} else {
> +				delta = clocksource_delta(wdnow, cs->wd_last_bogus, watchdog->mask);
> +				wd_lb = clocksource_cyc2ns(delta, watchdog->mult, watchdog->shift);
> +				if ((1 << cs->wd_bogus_shift) * wdi <= wd_lb)
> +					needwarn = true;
> +			}
> +			if (needwarn) {
> +				/* This can happen on busy systems, which can delay the watchdog. */
> +				pr_warn("timekeeping watchdog on CPU%d: Watchdog clocksource '%s' advanced an excessive %lld ns during %d-jiffy time interval (%lu additional), probable CPU overutilization, skipping watchdog check.\n", smp_processor_id(), watchdog->name, wd_nsec, WATCHDOG_INTERVAL, cs->wd_bogus_count - cs->wd_bogus_count_last);
> +				cs->wd_last_bogus = wdnow;
> +				if (cs->wd_bogus_shift < 10)
> +					cs->wd_bogus_shift++;
> +				cs->wd_bogus_count_last = cs->wd_bogus_count;

I don't think you need to keep 2 counters to store the number of skipped 
warnings. How about just use a single wd_bogus_skip_count like

If (needwarn) {
    pr_warn("... %lu additonal ...", ... cs->wd_bogus_skip_count);
    cs->wd_bogus_skip_count = 0:
} else {
    cs->wd_bogus_skip_count++;
}

Cheers,
Longman