Re: [PATCH] timekeeping: Add a lockdep override in tick_freeze().

[Peter Zijlstra]

On Fri, Apr 04, 2025 at 03:34:29PM +0200, Sebastian Andrzej Siewior wrote:
> tick_freeze() acquires a raw_spinlock_t (tick_freeze_lock). Later in the
> callchain (timekeeping_suspend() -> mc146818_avoid_UIP()) the RTC driver
> can acquire a spinlock_t which becomes a sleeping lock on PREEMPT_RT.
> Lockdep complains about this lock nesting.
> 
> Add a lockdep override for this special case and a comment explaining
> why it is okay.
> 
> Reported-by: Borislav Petkov <bp@xxxxxxxxx>
> Closes: https://lore.kernel.org/all/20250330113202.GAZ-krsjAnurOlTcp-@fat_crate.local/
> Reported-by: Chris Bainbridge <chris.bainbridge@xxxxxxxxx>
> Closes: https://lore.kernel.org/all/CAP-bSRZ0CWyZZsMtx046YV8L28LhY0fson2g4EqcwRAVN1Jk+Q@xxxxxxxxxxxxxx/
> Signed-off-by: Sebastian Andrzej Siewior <bigeasy@xxxxxxxxxxxxx>

This is of course horrible :-) But yes, probably the best one can do
given how things are.

Acked-by: Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>

> ---
>  kernel/time/tick-common.c | 20 ++++++++++++++++++++
>  1 file changed, 20 insertions(+)
> 
> diff --git a/kernel/time/tick-common.c b/kernel/time/tick-common.c
> index a47bcf71defcf..8fd8e2ee09fa1 100644
> --- a/kernel/time/tick-common.c
> +++ b/kernel/time/tick-common.c
> @@ -509,6 +509,7 @@ void tick_resume(void)
>  
>  #ifdef CONFIG_SUSPEND
>  static DEFINE_RAW_SPINLOCK(tick_freeze_lock);
> +static DEFINE_WAIT_OVERRIDE_MAP(tick_freeze_map, LD_WAIT_SLEEP);
>  static unsigned int tick_freeze_depth;
>  
>  /**
> @@ -528,9 +529,20 @@ void tick_freeze(void)
>  	if (tick_freeze_depth == num_online_cpus()) {
>  		trace_suspend_resume(TPS("timekeeping_freeze"),
>  				     smp_processor_id(), true);
> +		/*
> +		 * All other CPUs have their interrupts disabled and are
> +		 * suspended to idle. Other tasks have been frozen so there is
> +		 * no scheduling happening. This means that there is no
> +		 * concurrency in the system at this point. Therefore it is okay
> +		 * to acquire a sleeping lock on PREEMPT_RT, such as spinlock_t,
> +		 * because the lock can not be acquired and can not block.
> +		 * Inform lockdep about the situation.
> +		 */
> +		lock_map_acquire_try(&tick_freeze_map);
>  		system_state = SYSTEM_SUSPEND;
>  		sched_clock_suspend();
>  		timekeeping_suspend();
> +		lock_map_release(&tick_freeze_map);
>  	} else {
>  		tick_suspend_local();
>  	}
> @@ -552,8 +564,16 @@ void tick_unfreeze(void)
>  	raw_spin_lock(&tick_freeze_lock);
>  
>  	if (tick_freeze_depth == num_online_cpus()) {
> +		/*
> +		 * Similar to tick_freeze(). On resumption the first CPU may
> +		 * acquire uncontended sleeping locks while other CPUs block on
> +		 * tick_freeze_lock.
> +		 */
> +		lock_map_acquire_try(&tick_freeze_map);
>  		timekeeping_resume();
>  		sched_clock_resume();
> +		lock_map_release(&tick_freeze_map);
> +
>  		system_state = SYSTEM_RUNNING;
>  		trace_suspend_resume(TPS("timekeeping_freeze"),
>  				     smp_processor_id(), false);
> -- 
> 2.49.0
>