Re: [PATCH 1/4] generic-smp: remove single ipi fallback for smp_call_function_many()

[Oleg Nesterov <oleg@redhat.com>]

On 02/16, Oleg Nesterov wrote:
>
> > +	data->refs = cpumask_weight(&data->cpumask);
> > +
> > +	spin_lock_irqsave(&call_function.lock, flags);
> > +	call_function.counter++;
> > +	list_add_tail_rcu(&data->csd.list, &call_function.queue);
> > +	spin_unlock_irqrestore(&call_function.lock, flags);
>
> What if the initialization above leaks into the critical section?
>
> I mean, generic_smp_call_function_interrupt() running on another CPU
> can see the result of list_add_tail_rcu() and cpumask_and(data->cpumask)
> but not (say) "data->refs = ...".

I was wrong, sorry.

list_add_tail_rcu() implies wmb() and list_for_each_entry_rcu() does
smp_read_barrier_depends(), so another CPU must always see the fully
initialized entry.

Oleg.