Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Date:   Wed, 5 Jul 2023 14:45:00 +0200
From:   Frederic Weisbecker <frederic@kernel.org>
To:     Joel Fernandes <joel@joelfernandes.org>
Cc:     "Paul E . McKenney" <paulmck@kernel.org>,
        LKML <linux-kernel@vger.kernel.org>, rcu <rcu@vger.kernel.org>,
        Uladzislau Rezki <urezki@gmail.com>,
        Neeraj Upadhyay <quic_neeraju@quicinc.com>,
        Giovanni Gherdovich <ggherdovich@suse.cz>
Subject: Re: [PATCH 4/9] rcu: Introduce lazy queue's own qhimark
Message-ID: <ZKVlzOuQ0HyhSZ3L@localhost.localdomain>
References: <20230531101736.12981-1-frederic@kernel.org>
 <20230531101736.12981-5-frederic@kernel.org>
 <20230603012338.GA2795276@google.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=iso-8859-1
Content-Disposition: inline
Content-Transfer-Encoding: 8bit
In-Reply-To: <20230603012338.GA2795276@google.com>
Precedence: bulk

Le Sat, Jun 03, 2023 at 01:23:38AM +0000, Joel Fernandes a ?crit :
> On Wed, May 31, 2023 at 12:17:31PM +0200, Frederic Weisbecker wrote:
> > The lazy and the regular bypass queues share the same thresholds in
> > terms of number of callbacks after which a flush to the main list is
> > performed: 10 000 callbacks.
> > 
> > However lazy and regular bypass don't have the same purposes and neither
> > should their respective thresholds:
> > 
> > * The bypass queue stands for relieving the main queue in case of a
> >   callback storm. It makes sense to allow a high number of callbacks to
> >   pile up before flushing to the main queue, especially as the life
> >   cycle for this queue is very short (1 jiffy).
> 
> Sure.
> 
> > 
> > * The lazy queue aims to spare wake ups and reduce the number of grace
> >   periods. There it doesn't make sense to allow a huge number of
> >   callbacks before flushing so as not to introduce memory pressure,
> >   especially as the life cycle for this queue is very long (10
> >   seconds).
> 
> It does make sense as we have a shrinker, and it is better to avoid RCU
> disturbing the system unwantedly when there's lots of memory lying around.
> 
> > 
> > For those reasons, set the default threshold for the lazy queue to
> > 100.
> 
> I am OK with splitting the qhimark, but this lazy default is too low. In
> typical workloads on ChromeOS, we see 1000s of callback even when CPU
> utilization is low. So considering that, we would be flushing all the time.
> 
> Eventually I want the mm subsystem to tell us when flushing is needed so we
> could flush sooner at that time if really needed, but for now we have a
> shrinker so it should be OK. Is there a reason the shrinker does not work for
> you?

So you mean dynamically adapting the lazy qhimark on top of shrinker calls,
right? That would make sense indeed.

Thanks.