Re: [PATCH 1/3] jump_labels: fix concurrent static_key_enable/disable()

[Peter Zijlstra]

Thanks for doing these patches, I hadn't come around to them yet.

On Tue, Aug 01, 2017 at 05:24:04PM +0200, Paolo Bonzini wrote:
>  
> +void static_key_enable(struct static_key *key)
> +{
> +	STATIC_KEY_CHECK_USE();
> +	if (atomic_read(&key->enabled) > 0) {
> +		WARN_ON_ONCE(atomic_read(&key->enabled) != 1);
> +		return;
> +	}
> +
> +	cpus_read_lock();
> +	jump_label_lock();
> +	if (atomic_read(&key->enabled) == 0) {
> +		atomic_set(&key->enabled, -1);
> +		jump_label_update(key);

As per the previous discussion, should I do a patch adding barriers here
(or using atomic_set_release()) such that we close the window where a
concurrent inc/enable sees 1 but not all text changes?

> +		atomic_set(&key->enabled, 1);
> +	}
> +	jump_label_unlock();
> +	cpus_read_unlock();
> +}
> +EXPORT_SYMBOL_GPL(static_key_enable);