Hello, Hillf!
> Hi Ulad
>
> Good work with a nit.
>
Thank you :)
> On Mon, Oct 16, 2023 at 1:30 PM Uladzislau Rezki (Sony) <urezki@xxxxxxxxx> wrote:
> > +static void rcu_sr_normal_gp_init(void)
> > +{
> > + struct llist_node *llnode, *rcu;
> > + int ret;
> > +
> > + if (llist_empty(&sr.curr))
> > + return;
>
> This empty check erases the curr_tail race below instead of
> atomic_inc_return(&sr.active), because llist_add() will never return true
> after this check.
>
I use "active" counter to guarantee that a tail was updated in the
rcu_sr_normal_add_req(), i.e. the list might be not empty whereas the
tail updating might be in progress. llist_add() success and the task gets
preemted as an example.
Or i miss your point? If so, i appreciate if you clarify it in more
detail.
> > +
> > + /*
> > + * A waiting list of GP should be empty on this step,
> > + * since a GP-kthread, rcu_gp_init() -> gp_cleanup(),
> > + * rolls it over. If not, it is a BUG, warn a user.
> > + */
> > + WARN_ON_ONCE(!llist_empty(&sr.wait));
> > +
> > + /*
> > + * Obtain a tail of current active users. It is guaranteed
> > + * that if we are only one active user and the list is not
> > + * empty, the tail has already been updated.
> > + */
> > + ret = atomic_inc_return(&sr.active);
>
> Replacing atomic_inc_return() with smp_mb() cuts sr.active off.
>
But here we would like to know that we were only one user + not
empty list gurantees that a tail is ready.
Thank you for your comments!
--
Uladzislau Rezki
