22.12.2020 04:37, Frederic Weisbecker пишет:
> Entering RCU idle mode may cause a deferred wake up of an RCU NOCB_GP
> kthread (rcuog) to be serviced.
>
> Usually a wake up happening while running the idle task is spotted in
> one of the need_resched() checks carefully placed within the idle loop
> that can break to the scheduler.
>
> Unfortunately within cpuidle the call to rcu_idle_enter() is already
> beyond the last generic need_resched() check. Some drivers may perform
> their own checks like with mwait_idle_with_hints() but many others don't
> and we may halt the CPU with a resched request unhandled, leaving the
> task hanging.
>
> Fix this with performing a last minute need_resched() check after
> calling rcu_idle_enter().
>
> Reported-by: Paul E. McKenney <paulmck@xxxxxxxxxx>
> Fixes: 1098582a0f6c (sched,idle,rcu: Push rcu_idle deeper into the idle path)
> Cc: stable@xxxxxxxxxxxxxxx
> Cc: Daniel Lezcano <daniel.lezcano@xxxxxxxxxx>
> Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
> Cc: Rafael J. Wysocki <rafael.j.wysocki@xxxxxxxxx>
> Cc: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
> Cc: Ingo Molnar <mingo@xxxxxxxxxx>
> Signed-off-by: Frederic Weisbecker <frederic@xxxxxxxxxx>
> ---
> drivers/cpuidle/cpuidle.c | 33 +++++++++++++++++++++++++--------
> 1 file changed, 25 insertions(+), 8 deletions(-)
>
> diff --git a/drivers/cpuidle/cpuidle.c b/drivers/cpuidle/cpuidle.c
> index ef2ea1b12cd8..4cc1ba49ce05 100644
> --- a/drivers/cpuidle/cpuidle.c
> +++ b/drivers/cpuidle/cpuidle.c
> @@ -134,8 +134,8 @@ int cpuidle_find_deepest_state(struct cpuidle_driver *drv,
> }
>
> #ifdef CONFIG_SUSPEND
> -static void enter_s2idle_proper(struct cpuidle_driver *drv,
> - struct cpuidle_device *dev, int index)
> +static int enter_s2idle_proper(struct cpuidle_driver *drv,
> + struct cpuidle_device *dev, int index)
> {
> ktime_t time_start, time_end;
> struct cpuidle_state *target_state = &drv->states[index];
> @@ -151,7 +151,14 @@ static void enter_s2idle_proper(struct cpuidle_driver *drv,
> stop_critical_timings();
> if (!(target_state->flags & CPUIDLE_FLAG_RCU_IDLE))
> rcu_idle_enter();
> - target_state->enter_s2idle(dev, drv, index);
> + /*
> + * Last need_resched() check must come after rcu_idle_enter()
> + * which may wake up RCU internal tasks.
> + */
> + if (!need_resched())
> + target_state->enter_s2idle(dev, drv, index);
> + else
> + index = -EBUSY;
> if (WARN_ON_ONCE(!irqs_disabled()))
> local_irq_disable();
> if (!(target_state->flags & CPUIDLE_FLAG_RCU_IDLE))
> @@ -159,10 +166,13 @@ static void enter_s2idle_proper(struct cpuidle_driver *drv,
> tick_unfreeze();
> start_critical_timings();
>
> - time_end = ns_to_ktime(local_clock());
> + if (index > 0) {
index=0 is valid too
> + time_end = ns_to_ktime(local_clock());
> + dev->states_usage[index].s2idle_time += ktime_us_delta(time_end, time_start);
> + dev->states_usage[index].s2idle_usage++;
> + }
>
> - dev->states_usage[index].s2idle_time += ktime_us_delta(time_end, time_start);
> - dev->states_usage[index].s2idle_usage++;
> + return index;
> }
>
> /**
> @@ -184,7 +194,7 @@ int cpuidle_enter_s2idle(struct cpuidle_driver *drv, struct cpuidle_device *dev)
> */
> index = find_deepest_state(drv, dev, U64_MAX, 0, true);
> if (index > 0) {
> - enter_s2idle_proper(drv, dev, index);
> + index = enter_s2idle_proper(drv, dev, index);
> local_irq_enable();
> }
> return index;
> @@ -234,7 +244,14 @@ int cpuidle_enter_state(struct cpuidle_device *dev, struct cpuidle_driver *drv,
> stop_critical_timings();
> if (!(target_state->flags & CPUIDLE_FLAG_RCU_IDLE))
> rcu_idle_enter();
> - entered_state = target_state->enter(dev, drv, index);
> + /*
> + * Last need_resched() check must come after rcu_idle_enter()
> + * which may wake up RCU internal tasks.
> + */
> + if (!need_resched())
> + entered_state = target_state->enter(dev, drv, index);
> + else
> + entered_state = -EBUSY;
> if (!(target_state->flags & CPUIDLE_FLAG_RCU_IDLE))
> rcu_idle_exit();
> start_critical_timings();
>
This patch causes a hardlock on NVIDIA Tegra using today's linux-next.
Disabling coupled idling state helps. Please fix thanks in advance.
