Hello,
On Thu, Sep 20, 2018 at 06:18:21PM +0800, Jianchao Wang wrote:
> -static inline void percpu_ref_get_many(struct percpu_ref *ref, unsigned long nr)
> +static inline void __percpu_ref_get_many(struct percpu_ref *ref, unsigned long nr)
> {
> unsigned long __percpu *percpu_count;
>
> - rcu_read_lock_sched();
So, if we're gonna do this (please read below tho), please add the
matching assertion
> if (__ref_is_percpu(ref, &percpu_count))
> this_cpu_add(*percpu_count, nr);
> else
> atomic_long_add(nr, &ref->count);
> +}
>
> +/**
> + * percpu_ref_get_many - increment a percpu refcount
> + * @ref: percpu_ref to get
> + * @nr: number of references to get
> + *
> + * Analogous to atomic_long_add().
> + *
> + * This function is safe to call as long as @ref is between init and exit.
> + */
> +static inline void percpu_ref_get_many(struct percpu_ref *ref, unsigned long nr)
> +{
> + rcu_read_lock_sched();
> + __percpu_ref_get_many(ref, nr);
> rcu_read_unlock_sched();
> }
And add the matching variant for get/put with and without _many.
Ming, so, if we make locking explicit like above, I think it should be
fine to share the locking. However, please note that percpu_ref and
blk_mq are using different types of RCU, at least for now, and I'm not
really sure that unifying that and taking out one rcu read lock/unlock
is a meaningful optimization.
Let's please first do something straight-forward. If somebody can
show that this actually impacts performance, we can optimize it but
right now all these seem premature to me.
Thanks.
--
tejun
