Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.37 as permitted sender) client-ip=23.128.96.37;
From:   Uladzislau Rezki <urezki@gmail.com>
Date:   Thu, 26 Oct 2023 15:00:25 +0200
To:     Frederic Weisbecker <frederic@kernel.org>
Cc:     "Uladzislau Rezki (Sony)" <urezki@gmail.com>,
        "Paul E . McKenney" <paulmck@kernel.org>,
        RCU <rcu@vger.kernel.org>,
        Neeraj upadhyay <Neeraj.Upadhyay@amd.com>,
        Boqun Feng <boqun.feng@gmail.com>,
        Hillf Danton <hdanton@sina.com>,
        Joel Fernandes <joel@joelfernandes.org>,
        LKML <linux-kernel@vger.kernel.org>,
        Oleksiy Avramchenko <oleksiy.avramchenko@sony.com>
Subject: Re: [PATCH 1/3] rcu: Reduce synchronize_rcu() waiting time
Message-ID: <ZTpi6diP4h84PtWE@pc636>
References: <20231025140915.590390-1-urezki@gmail.com>
 <20231025140915.590390-2-urezki@gmail.com>
 <ZTkwl0bzTTCy8g5N@localhost.localdomain>
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Disposition: inline
Content-Transfer-Encoding: 8bit
In-Reply-To: <ZTkwl0bzTTCy8g5N@localhost.localdomain>
Precedence: bulk

On Wed, Oct 25, 2023 at 05:13:27PM +0200, Frederic Weisbecker wrote:
> Le Wed, Oct 25, 2023 at 04:09:13PM +0200, Uladzislau Rezki (Sony) a écrit :
> > +/*
> > + * Helper function for rcu_gp_cleanup().
> > + */
> > +static void rcu_sr_normal_gp_cleanup(void)
> > +{
> > +	struct llist_node *head, *tail, *pos;
> > +	int i = 0;
> > +
> > +	tail = READ_ONCE(sr.wait_tail);
> > +	head = llist_del_all(&sr.wait);
> 
> This could be llist_empty() first to do a quick
> cheap check. And then __llist_del_all() here because
> it appears nothing else than gp kthread can touch sr.wait.
> 
No problem i can fix it. Initially i had a check first!

> > +
> > +	llist_for_each_safe(pos, head, head) {
> 
> Two times head intended here? There should be some
> temporary storage in the middle.
> 
Yes. It is intentially done. The head is updated, i.e. shifted to a next,
because we directly process users from a GP. The number is limited to 5
all the rest is deferred.

> > +		rcu_sr_normal_complete(pos);
> > +
> > +		if (++i == MAX_SR_WAKE_FROM_GP) {
> > +			/* If last, process it also. */
> > +			if (head && !head->next)
> > +				continue;
> > +			break;
> > +		}
> > +	}
> > +
> > +	if (head) {
> > +		/* Can be not empty. */
> > +		llist_add_batch(head, tail, &sr.done);
> > +		queue_work(system_highpri_wq, &sr_normal_gp_cleanup);
> 
> So you can have:
> 
> * Queue to sr.curr is atomic fully ordered
> * Check and move from sr.curr to sr.wait is atomic fully ordered
> * Check from sr.wait can have a quick unatomic unordered
>   llist_empty() check. Then extract unatomic unordered as well.
> * If too many, move atomic/ordered to sr.done.
> 
> Am I missing something?
>
If too many move to done and kick the helper. The sr.wait can not
be touched until the rcu_sr_normal_gp_cleanup() is completed, i.e.:

<snip>
GP-kthread(same and one task context):
    rcu_sr_normal_gp_cleanup();
    wait for a grace period;
    rcu_sr_normal_gp_cleanup();
<snip>

Am i missing your point?

--
Uladzislau Rezki