Re: [RFC PATCH v4 1/9] CPU hotplug: Provide APIs to prevent CPU offline from atomic context

[Oleg Nesterov <oleg@redhat.com>]

On 12/18, Srivatsa S. Bhat wrote:
>
> So now that we can't avoid disabling and enabling interrupts,

Still I think it would be better to not use local_irq_save/restore
directly. And,

> I was
> wondering if we could exploit this to avoid the smp_mb()..
>
> Maybe this is a stupid question, but I'll shoot it anyway...
> Does local_irq_disable()/enable provide any ordering guarantees by any chance?

Oh, I do not know.

But please look at the comment above prepare_to_wait(). It assumes
that even spin_unlock_irqrestore() is not the full barrier.

In any case. get_online_cpus_atomic() has to use irq_restore, not
irq_enable. And _restore does nothing "special" if irqs were already
disabled, so I think we can't rely on sti.

> I tried thinking about other ways to avoid that smp_mb() in the reader,

Just in case, I think there is no way to avoid mb() in _get (although
perhaps it can be "implicit").

The writer changes cpu_online_mask and drops the lock. We need to ensure
that the reader sees the change in cpu_online_mask after _get returns.

> but was unsuccessful. So if the above assumption is wrong, I guess we'll
> just have to go with the version that uses synchronize_sched() at the
> writer-side.

In this case we can also convert get_online_cpus() to use percpu_rwsem
and avoid mutex_lock(&cpu_hotplug.lock), but this is minor I guess.
I do not think get_online_cpus() is called too often.

Oleg.