Re: [PATCH 1/9] rcu: Assume IRQS disabled from rcu_report_dead()

[Joel Fernandes <joel@joelfernandes.org>]

On Wed, May 31, 2023 at 12:17:28PM +0200, Frederic Weisbecker wrote:
> rcu_report_dead() is the last RCU word from the CPU down through the
> hotplug path. It is called in the idle loop right before the CPU shuts
> down for good. Because it removes the CPU from the grace period state
> machine and reports an ultimate quiescent state if necessary, no further
> use of RCU is allowed. Therefore it is expected that IRQs are disabled
> upon calling this function and are not to be re-enabled again until the
> CPU shuts down.
> 
> Remove the IRQs disablement from that function and verify instead that
> it is actually called with IRQs disabled as it is expected at that
> special point in the idle path.
> 
> Signed-off-by: Frederic Weisbecker <frederic@kernel.org>
> ---
>  kernel/rcu/tree.c | 10 ++++++----
>  1 file changed, 6 insertions(+), 4 deletions(-)
> 
> diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> index fae9b4e29c93..bc4e7c9b51cb 100644
> --- a/kernel/rcu/tree.c
> +++ b/kernel/rcu/tree.c
> @@ -4476,11 +4476,16 @@ void rcu_cpu_starting(unsigned int cpu)
>   */
>  void rcu_report_dead(unsigned int cpu)
>  {
> -	unsigned long flags, seq_flags;
> +	unsigned long flags;
>  	unsigned long mask;
>  	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
>  	struct rcu_node *rnp = rdp->mynode;  /* Outgoing CPU's rdp & rnp. */
>  
> +	/*
> +	 * IRQS must be disabled from now on and until the CPU dies, or an interrupt
> +	 * may introduce a new READ-side while it is actually off the QS masks.
> +	 */
> +	lockdep_assert_irqs_disabled();
>  	// Do any dangling deferred wakeups.
>  	do_nocb_deferred_wakeup(rdp);
>  
> @@ -4488,7 +4493,6 @@ void rcu_report_dead(unsigned int cpu)
>  
>  	/* Remove outgoing CPU from mask in the leaf rcu_node structure. */
>  	mask = rdp->grpmask;
> -	local_irq_save(seq_flags);

True, IRQs should be disabled here. The idle loop disables irqs before
calling cpuhp_report_idle_dead() which calls rcu_report_dead().

I was curious about this path called from cpu_die_early() in ARM, in which
case it is an existing bug if it did not already disable interrupts. So your
lockdep check is a good thing in that regard.

For this patch:
Reviewed-by: Joel Fernandes (Google) <joel@joelfernandes.org>

thanks,

 - Joel


>  	arch_spin_lock(&rcu_state.ofl_lock);
>  	raw_spin_lock_irqsave_rcu_node(rnp, flags); /* Enforce GP memory-order guarantee. */
>  	rdp->rcu_ofl_gp_seq = READ_ONCE(rcu_state.gp_seq);
> @@ -4502,8 +4506,6 @@ void rcu_report_dead(unsigned int cpu)
>  	WRITE_ONCE(rnp->qsmaskinitnext, rnp->qsmaskinitnext & ~mask);
>  	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
>  	arch_spin_unlock(&rcu_state.ofl_lock);
> -	local_irq_restore(seq_flags);
> -
>  	rdp->cpu_started = false;
>  }
>  
> -- 
> 2.40.1
>