Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Date:   Thu, 5 Mar 2020 17:25:57 -0500
From:   Joel Fernandes <joel@joelfernandes.org>
To:     linux-kernel@vger.kernel.org
Cc:     urezki@gmail.com, Davidlohr Bueso <dave@stgolabs.net>,
        Josh Triplett <josh@joshtriplett.org>,
        Lai Jiangshan <jiangshanlai@gmail.com>,
        Mathieu Desnoyers <mathieu.desnoyers@efficios.com>,
        "Paul E. McKenney" <paulmck@kernel.org>, rcu@vger.kernel.org,
        Steven Rostedt <rostedt@goodmis.org>
Subject: Re: [PATCH linus/master 2/2] rcu/tree: Add a shrinker to prevent OOM
 due to kfree_rcu() batching
Message-ID: <20200305222557.GC66450@google.com>
References: <20200305221323.66051-1-joel@joelfernandes.org>
 <20200305221323.66051-2-joel@joelfernandes.org>
 <20200305221753.GA66450@google.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20200305221753.GA66450@google.com>
User-Agent: Mutt/1.12.2 (2019-09-21)
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Thu, Mar 05, 2020 at 05:17:53PM -0500, Joel Fernandes wrote:
> On Thu, Mar 05, 2020 at 05:13:23PM -0500, Joel Fernandes (Google) wrote:
> > To reduce grace periods and improve kfree() performance, we have done
> > batching recently dramatically bringing down the number of grace periods
> > while giving us the ability to use kfree_bulk() for efficient kfree'ing.
> > 
> > However, this has increased the likelihood of OOM condition under heavy
> > kfree_rcu() flood on small memory systems. This patch introduces a
> > shrinker which starts grace periods right away if the system is under
> > memory pressure due to existence of objects that have still not started
> > a grace period.
> > 
> > With this patch, I do not observe an OOM anymore on a system with 512MB
> > RAM and 8 CPUs, with the following rcuperf options:
> > 
> > rcuperf.kfree_loops=20000 rcuperf.kfree_alloc_num=8000
> > rcuperf.kfree_rcu_test=1 rcuperf.kfree_mult=2
> 
> Paul,
> I may have to rebase this patch on top of Vlad's kfree_bulk() work. But let
> us discuss patch and I can rebase it and repost it once patch looks Ok to
> you. (The kfree_bulk() work should not affect the patch).

BTW, we can also use the scheme in the future to keep garbage uncollected
until memory pressure. That way you defer grace periods for longer similar to
the paper [1], until the MM layer thinks the party is over. For one, I am not
too confident about the shrinker's ability to handle transient memory spikes.
If I remember, the shrinker is best-effort.

But one step at a time :)

thanks,

 - Joel

[1] https://dl.acm.org/doi/10.1145/3190508.3190522