[PATCH] clocksource: Warn if too many missing ticks are detected

[PATCH] clocksource: Warn if too many missing ticks are detected

[Waiman Long]

The clocksource watchdog, when running, is scheduled on all the CPUs in
the system sequentially on a round-robin fashion with a period of 0.5s.
A bug in the 4.18 kernel is causing missing ticks when nohz_full
is specified. Under some circumstances, this causes the watchdog to
incorrectly state that the TSC is unstable because of counter overflow
in the hpet watchdog clock source after a few minutes delay.

That particular bug is fixed by the 4.19 commit 7059b36636beab ("sched:
idle: Avoid retaining the tick when it has been stopped"). To make it
easier to catch this kind of bug in the future, a check is added to see
if there is too much delay in the watchdog invocation and print a
warning once if it happens.

Signed-off-by: Waiman Long <longman@redhat.com>
---
 kernel/time/clocksource.c | 13 +++++++++++++
 1 file changed, 13 insertions(+)

diff --git a/kernel/time/clocksource.c b/kernel/time/clocksource.c
index 0e6e97a..2ea5db0 100644
--- a/kernel/time/clocksource.c
+++ b/kernel/time/clocksource.c
@@ -140,6 +140,7 @@ static void inline clocksource_watchdog_unlock(unsigned long *flags)
  * Interval: 0.5sec Threshold: 0.0625s
  */
 #define WATCHDOG_INTERVAL (HZ >> 1)
+#define WATCHDOG_INTERNVAL_NS (NSEC_PER_SEC >> 1)
 #define WATCHDOG_THRESHOLD (NSEC_PER_SEC >> 4)
 
 static void clocksource_watchdog_work(struct work_struct *work)
@@ -242,6 +243,18 @@ static void clocksource_watchdog(struct timer_list *unused)
 		wd_nsec = clocksource_cyc2ns(delta, watchdog->mult,
 					     watchdog->shift);
 
+		/*
+		 * When the timer tick is incorrectly stopped on a CPU with
+		 * pending events, for example, it is possible that the
+		 * clocksource watchdog will stop running for a sufficiently
+		 * long enough time to cause overflow in the delta
+		 * computation leading to incorrect report of unstable clock
+		 * source. So print a warning if there is unusually large
+		 * delay (> 0.5s) in the invocation of the watchdog. That
+		 * can indicate a hidden bug in the timer tick code.
+		 */
+		WARN_ON_ONCE(!wd_nsec || wd_nsec > 2*WATCHDOG_INTERNVAL_NS);
+
 		delta = clocksource_delta(csnow, cs->cs_last, cs->mask);
 		cs_nsec = clocksource_cyc2ns(delta, cs->mult, cs->shift);
 		wdlast = cs->wd_last; /* save these in case we print them */
-- 
1.8.3.1

Re: [PATCH] clocksource: Warn if too many missing ticks are detected

[Thomas Gleixner]

Waiman,

On Tue, 18 Sep 2018, Waiman Long wrote:

> The clocksource watchdog, when running, is scheduled on all the CPUs in
> the system sequentially on a round-robin fashion with a period of 0.5s.
> A bug in the 4.18 kernel is causing missing ticks when nohz_full
> is specified. Under some circumstances, this causes the watchdog to
> incorrectly state that the TSC is unstable because of counter overflow
> in the hpet watchdog clock source after a few minutes delay.
> 
> That particular bug is fixed by the 4.19 commit 7059b36636beab ("sched:
> idle: Avoid retaining the tick when it has been stopped"). To make it
> easier to catch this kind of bug in the future, a check is added to see
> if there is too much delay in the watchdog invocation and print a
> warning once if it happens.

I like the idea.

> 
> Signed-off-by: Waiman Long <longman@redhat.com>
> ---
>  kernel/time/clocksource.c | 13 +++++++++++++
>  1 file changed, 13 insertions(+)
> 
> diff --git a/kernel/time/clocksource.c b/kernel/time/clocksource.c
> index 0e6e97a..2ea5db0 100644
> --- a/kernel/time/clocksource.c
> +++ b/kernel/time/clocksource.c
> @@ -140,6 +140,7 @@ static void inline clocksource_watchdog_unlock(unsigned long *flags)
>   * Interval: 0.5sec Threshold: 0.0625s
>   */
>  #define WATCHDOG_INTERVAL (HZ >> 1)
> +#define WATCHDOG_INTERNVAL_NS (NSEC_PER_SEC >> 1)
>  #define WATCHDOG_THRESHOLD (NSEC_PER_SEC >> 4)
>  
>  static void clocksource_watchdog_work(struct work_struct *work)
> @@ -242,6 +243,18 @@ static void clocksource_watchdog(struct timer_list *unused)
>  		wd_nsec = clocksource_cyc2ns(delta, watchdog->mult,
>  					     watchdog->shift);
>  
> +		/*
> +		 * When the timer tick is incorrectly stopped on a CPU with
> +		 * pending events, for example, it is possible that the
> +		 * clocksource watchdog will stop running for a sufficiently
> +		 * long enough time to cause overflow in the delta
> +		 * computation leading to incorrect report of unstable clock
> +		 * source. So print a warning if there is unusually large
> +		 * delay (> 0.5s) in the invocation of the watchdog. That
> +		 * can indicate a hidden bug in the timer tick code.
> +		 */
> +		WARN_ON_ONCE(!wd_nsec || wd_nsec > 2*WATCHDOG_INTERNVAL_NS);

But this is using the watchdog delta to check. If that wrapped the
detection is broken.

I'd rather use watchdog_timer.expires and check against jiffies. That tells
you how late the timer callback actually is and does not suffer any
wraparound issues.

Thanks,

	tglx

Re: [PATCH] clocksource: Warn if too many missing ticks are detected

[Waiman Long]

On 09/18/2018 05:07 PM, Thomas Gleixner wrote:
> Waiman,
>
> On Tue, 18 Sep 2018, Waiman Long wrote:
>
>> The clocksource watchdog, when running, is scheduled on all the CPUs in
>> the system sequentially on a round-robin fashion with a period of 0.5s.
>> A bug in the 4.18 kernel is causing missing ticks when nohz_full
>> is specified. Under some circumstances, this causes the watchdog to
>> incorrectly state that the TSC is unstable because of counter overflow
>> in the hpet watchdog clock source after a few minutes delay.
>>
>> That particular bug is fixed by the 4.19 commit 7059b36636beab ("sched:
>> idle: Avoid retaining the tick when it has been stopped"). To make it
>> easier to catch this kind of bug in the future, a check is added to see
>> if there is too much delay in the watchdog invocation and print a
>> warning once if it happens.
> I like the idea.
>
>> Signed-off-by: Waiman Long <longman@redhat.com>
>> ---
>>  kernel/time/clocksource.c | 13 +++++++++++++
>>  1 file changed, 13 insertions(+)
>>
>> diff --git a/kernel/time/clocksource.c b/kernel/time/clocksource.c
>> index 0e6e97a..2ea5db0 100644
>> --- a/kernel/time/clocksource.c
>> +++ b/kernel/time/clocksource.c
>> @@ -140,6 +140,7 @@ static void inline clocksource_watchdog_unlock(unsigned long *flags)
>>   * Interval: 0.5sec Threshold: 0.0625s
>>   */
>>  #define WATCHDOG_INTERVAL (HZ >> 1)
>> +#define WATCHDOG_INTERNVAL_NS (NSEC_PER_SEC >> 1)
>>  #define WATCHDOG_THRESHOLD (NSEC_PER_SEC >> 4)
>>  
>>  static void clocksource_watchdog_work(struct work_struct *work)
>> @@ -242,6 +243,18 @@ static void clocksource_watchdog(struct timer_list *unused)
>>  		wd_nsec = clocksource_cyc2ns(delta, watchdog->mult,
>>  					     watchdog->shift);
>>  
>> +		/*
>> +		 * When the timer tick is incorrectly stopped on a CPU with
>> +		 * pending events, for example, it is possible that the
>> +		 * clocksource watchdog will stop running for a sufficiently
>> +		 * long enough time to cause overflow in the delta
>> +		 * computation leading to incorrect report of unstable clock
>> +		 * source. So print a warning if there is unusually large
>> +		 * delay (> 0.5s) in the invocation of the watchdog. That
>> +		 * can indicate a hidden bug in the timer tick code.
>> +		 */
>> +		WARN_ON_ONCE(!wd_nsec || wd_nsec > 2*WATCHDOG_INTERNVAL_NS);
> But this is using the watchdog delta to check. If that wrapped the
> detection is broken.
>
> I'd rather use watchdog_timer.expires and check against jiffies. That tells
> you how late the timer callback actually is and does not suffer any
> wraparound issues.

The clocksource_delta() function will deal with wrap-around in the
counter value. It is only when the counter advances more than 0x80000000
for 32-bit hpet counter mask that a value of 0 will be returned. That is
why I have a !wd_nsec check there. There is a small chance when the
warparound is just within the 1 second window that the test fails. In
this case, the following kind of warning will certainly be triggered:

[  578.890937] clocksource: timekeeping watchdog on CPU21: Marking
clocksource 'tsc' as unstable because the skew is too large:
[  578.890938] clocksource:                       'hpet' wd_now:
ee332105 wd_last: 544f80e7 mask: ffffffff
[  578.890939] clocksource:                       'tsc' cs_now:
4b6e6ccb5d609 cs_last: 4b679a469d09e mask: ffffffffffffffff
[  578.890954] tsc: Marking TSC unstable due to clocksource watchdog
[  578.890963] TSC found unstable after boot, most likely due to broken
BIOS. Use 'tsc=unstable'.
[  578.890965] sched_clock: Marking unstable (578920214163,
-28725675)<-(579047174801, -156217937)
[  578.891056] clocksource: Switched to clocksource hpet

Another reason that I used wd_nsec is because the data has already been
computed.

I am perfectly fine to use the watchdog_timer.expires as suggested, though.

Cheers,
Longman

watchdog_timer.expires

Re: [PATCH] clocksource: Warn if too many missing ticks are detected

[Thomas Gleixner]

On Tue, 18 Sep 2018, Waiman Long wrote:
> On 09/18/2018 05:07 PM, Thomas Gleixner wrote:
> > On Tue, 18 Sep 2018, Waiman Long wrote:
> >> --- a/kernel/time/clocksource.c
> >> +++ b/kernel/time/clocksource.c
> >> @@ -140,6 +140,7 @@ static void inline clocksource_watchdog_unlock(unsigned long *flags)
> >>   * Interval: 0.5sec Threshold: 0.0625s
> >>   */
> >>  #define WATCHDOG_INTERVAL (HZ >> 1)
> >> +#define WATCHDOG_INTERNVAL_NS (NSEC_PER_SEC >> 1)
> >>  #define WATCHDOG_THRESHOLD (NSEC_PER_SEC >> 4)
> >>  
> >>  static void clocksource_watchdog_work(struct work_struct *work)
> >> @@ -242,6 +243,18 @@ static void clocksource_watchdog(struct timer_list *unused)
> >>  		wd_nsec = clocksource_cyc2ns(delta, watchdog->mult,
> >>  					     watchdog->shift);
> >>  
> >> +		/*
> >> +		 * When the timer tick is incorrectly stopped on a CPU with
> >> +		 * pending events, for example, it is possible that the
> >> +		 * clocksource watchdog will stop running for a sufficiently
> >> +		 * long enough time to cause overflow in the delta
> >> +		 * computation leading to incorrect report of unstable clock
> >> +		 * source. So print a warning if there is unusually large
> >> +		 * delay (> 0.5s) in the invocation of the watchdog. That
> >> +		 * can indicate a hidden bug in the timer tick code.
> >> +		 */
> >> +		WARN_ON_ONCE(!wd_nsec || wd_nsec > 2*WATCHDOG_INTERNVAL_NS);
> > But this is using the watchdog delta to check. If that wrapped the
> > detection is broken.
> >
> > I'd rather use watchdog_timer.expires and check against jiffies. That tells
> > you how late the timer callback actually is and does not suffer any
> > wraparound issues.
> 
> The clocksource_delta() function will deal with wrap-around in the
> counter value. It is only when the counter advances more than 0x80000000
> for 32-bit hpet counter mask that a value of 0 will be returned. That is
> why I have a !wd_nsec check there. There is a small chance when the
> warparound is just within the 1 second window that the test fails. In
> this case, the following kind of warning will certainly be triggered:

Ok.

> [  578.890937] clocksource: timekeeping watchdog on CPU21: Marking
> clocksource 'tsc' as unstable because the skew is too large:
> [  578.890938] clocksource:                       'hpet' wd_now:
> ee332105 wd_last: 544f80e7 mask: ffffffff
> [  578.890939] clocksource:                       'tsc' cs_now:
> 4b6e6ccb5d609 cs_last: 4b679a469d09e mask: ffffffffffffffff
> [  578.890954] tsc: Marking TSC unstable due to clocksource watchdog
> [  578.890963] TSC found unstable after boot, most likely due to broken
> BIOS. Use 'tsc=unstable'.
> [  578.890965] sched_clock: Marking unstable (578920214163,
> -28725675)<-(579047174801, -156217937)
> [  578.891056] clocksource: Switched to clocksource hpet
> 
> Another reason that I used wd_nsec is because the data has already been
> computed.

Sure, but that's hardly a hotpath.

> I am perfectly fine to use the watchdog_timer.expires as suggested, though.

I really prefer that because it's just more robust and obvious. No head
scratching about correctness required.

Thanks,

	tglx