[PATCH v6 08/46] CPU hotplug: Provide APIs to prevent CPU offline from atomic context

[Michel Lespinasse]

On Mon, Feb 18, 2013 at 8:39 PM, Srivatsa S. Bhat
<srivatsa.bhat at linux.vnet.ibm.com> wrote:
> Some important design requirements and considerations:
> -----------------------------------------------------
>
> 1. Scalable synchronization at the reader-side, especially in the fast-path
>
>    Any synchronization at the atomic hotplug readers side must be highly
>    scalable - avoid global single-holder locks/counters etc. Because, these
>    paths currently use the extremely fast preempt_disable(); our replacement
>    to preempt_disable() should not become ridiculously costly and also should
>    not serialize the readers among themselves needlessly.
>
>    At a minimum, the new APIs must be extremely fast at the reader side
>    atleast in the fast-path, when no CPU offline writers are active.
>
> 2. preempt_disable() was recursive. The replacement should also be recursive.
>
> 3. No (new) lock-ordering restrictions
>
>    preempt_disable() was super-flexible. It didn't impose any ordering
>    restrictions or rules for nesting. Our replacement should also be equally
>    flexible and usable.
>
> 4. No deadlock possibilities
>
>    Regular per-cpu locking is not the way to go if we want to have relaxed
>    rules for lock-ordering. Because, we can end up in circular-locking
>    dependencies as explained in https://lkml.org/lkml/2012/12/6/290
>
>    So, avoid the usual per-cpu locking schemes (per-cpu locks/per-cpu atomic
>    counters with spin-on-contention etc) as much as possible, to avoid
>    numerous deadlock possibilities from creeping in.
>
>
> Implementation of the design:
> ----------------------------
>
> We use per-CPU reader-writer locks for synchronization because:
>
>   a. They are quite fast and scalable in the fast-path (when no writers are
>      active), since they use fast per-cpu counters in those paths.
>
>   b. They are recursive at the reader side.
>
>   c. They provide a good amount of safety against deadlocks; they don't
>      spring new deadlock possibilities on us from out of nowhere. As a
>      result, they have relaxed locking rules and are quite flexible, and
>      thus are best suited for replacing usages of preempt_disable() or
>      local_irq_disable() at the reader side.
>
> Together, these satisfy all the requirements mentioned above.

Thanks for this detailed design explanation.

> +/*
> + * Invoked by atomic hotplug reader (a task which wants to prevent
> + * CPU offline, but which can't afford to sleep), to prevent CPUs from
> + * going offline. So, you can call this function from atomic contexts
> + * (including interrupt handlers).
> + *
> + * Note: This does NOT prevent CPUs from coming online! It only prevents
> + * CPUs from going offline.
> + *
> + * You can call this function recursively.
> + *
> + * Returns with preemption disabled (but interrupts remain as they are;
> + * they are not disabled).
> + */
> +void get_online_cpus_atomic(void)
> +{
> +       percpu_read_lock_irqsafe(&hotplug_pcpu_rwlock);
> +}
> +EXPORT_SYMBOL_GPL(get_online_cpus_atomic);
> +
> +void put_online_cpus_atomic(void)
> +{
> +       percpu_read_unlock_irqsafe(&hotplug_pcpu_rwlock);
> +}
> +EXPORT_SYMBOL_GPL(put_online_cpus_atomic);

So, you made it clear why you want a recursive read side here.

I am wondering though, if you could take care of recursive uses in
get/put_online_cpus_atomic() instead of doing it as a property of your
rwlock:

get_online_cpus_atomic()
{
    unsigned long flags;
    local_irq_save(flags);
    if (this_cpu_inc_return(hotplug_recusion_count) == 1)
        percpu_read_lock_irqsafe(&hotplug_pcpu_rwlock);
    local_irq_restore(flags);
}

Once again, the idea there is to avoid baking the reader side
recursive properties into your rwlock itself, so that it won't be
impossible to implement reader/writer fairness into your rwlock in the
future (which may be not be very important for the hotplug use, but
could be when other uses get introduced).

-- 
Michel "Walken" Lespinasse
A program is never fully debugged until the last user dies.