There have been reports of the watchdog marking clocksources unstable on
machines with 8 NUMA nodes:

  clocksource: timekeeping watchdog on CPU373:
  Marking clocksource 'tsc' as unstable because the skew is too large:
  clocksource:   'hpet' wd_nsec: 14523447520
  clocksource:   'tsc'  cs_nsec: 14524115132

The measured clocksource skew - the absolute difference between cs_nsec
and wd_nsec - was 668 microseconds:

  cs_nsec - wd_nsec = 14524115132 - 14523447520 = 667612

The kernel used 200 microseconds for the uncertainty_margin of both the
clocksource and watchdog, resulting in a threshold of 400 microseconds (the
md variable). Both the cs_nsec and the wd_nsec value indicate that the
readout interval was circa 14.5 seconds.  The observed behaviour is that
watchdog checks failed for large readout intervals on 8 NUMA node
machines. This indicates that the size of the skew was directly proportinal
to the length of the readout interval on those machines. The measured
clocksource skew, 668 microseconds, was evaluated against a threshold (the
md variable) that is suited for readout intervals of roughly
WATCHDOG_INTERVAL, i.e. HZ >> 1, which is 0.5 second.

The intention of 2e27e793 ("clocksource: Reduce clocksource-skew
threshold") was to tighten the threshold for evaluating skew and set the
lower bound for the uncertainty_margin of clocksources to twice
WATCHDOG_MAX_SKEW. Later in c37e85c1 ("clocksource: Loosen clocksource
watchdog constraints"), the WATCHDOG_MAX_SKEW constant was increased to
125 microseconds to fit the limit of NTP, which is able to use a
clocksource that suffers from up to 500 microseconds of skew per second.
Both the TSC and the HPET use default uncertainty_margin. When the
readout interval gets stretched the default uncertainty_margin is no
longer a suitable lower bound for evaluating skew - it imposes a limit
that is far stricter than the skew with which NTP can deal.

The root causes of the skew being directly proportinal to the length of
the readout interval are:

  * the inaccuracy of the shift/mult pairs of clocksources and the watchdog
  * the conversion to nanoseconds is imprecise for large readout intervals

Prevent this by skipping the current watchdog check if the readout
interval exceeds 2 * WATCHDOG_INTERVAL. Considering the maximum readout
interval of 2 * WATCHDOG_INTERVAL, the current default uncertainty margin
(of the TSC and HPET) corresponds to a limit on clocksource skew of 250
ppm (microseconds of skew per second).  To keep the limit imposed by NTP
(500 microseconds of skew per second) for all possible readout intervals,
the margins would have to be scaled so that the threshold value is
proportional to the length of the actual readout interval.

As for why the readout interval may get stretched: Since the watchdog is
executed in softirq context the expiration of the watchdog timer can get
severely delayed on account of a ksoftirqd thread not getting to run in a
timely manner. Surely, a system with such belated softirq execution is not
working well and the scheduling issue should be looked into but the
clocksource watchdog should be able to deal with it accordingly.

Fixes: 2e27e793 ("clocksource: Reduce clocksource-skew threshold")
Suggested-by: Feng Tang <feng.tang@intel.com>
Signed-off-by: Jiri Wiesner <jwiesner@suse.de>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Tested-by: Paul E. McKenney <paulmck@kernel.org>
Reviewed-by: Feng Tang <feng.tang@intel.com>
Cc: stable@vger.kernel.org
Link: https://lore.kernel.org/r/20240122172350.GA740@incl