When the ladder governor sees the CPUIDLE_FLAG_TIME_INVALID flag,
it unconditionally causes a state promotion by setting last_residency
to a number higher than the state's promotion_time:

last_residency = last_state->threshold.promotion_time + 1

It does this for fear that cpuidle_get_last_residency()
will be in-accurate, because cpuidle_enter_state() invoked
a state with CPUIDLE_FLAG_TIME_INVALID.

But the only state with CPUIDLE_FLAG_TIME_INVALID is
acpi_safe_halt(), which may return well after its actual
idle duration because it enables interrupts, so cpuidle_enter_state()
also measures interrupt service time.

So what?  In ladder, a huge invalid last_residency has exactly
the same effect as the current code -- it unconditionally
causes a state promotion.

In the case where the idle residency plus measured interrupt
handling time is less than the state's demotion_time -- we should
use that timestamp to give ladder a chance to demote, rather than
unconditionally promoting.

This can be done by simply ignoring the CPUIDLE_FLAG_TIME_INVALID,
and using the "invalid" time, as it is either equal to what we are
doing today, or better.
Signed-off-by: Len Brown <len.brown@intel.com>
Acked-by: Daniel Lezcano <daniel.lezcano@linaro.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>