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, 12 Jul 2023 13:48:15 -0700
From:   "Paul E. McKenney" <paulmck@kernel.org>
To:     Frederic Weisbecker <frederic@kernel.org>
Cc:     kernel test robot <oliver.sang@intel.com>, oe-lkp@lists.linux.dev,
        lkp@intel.com, linux-kernel@vger.kernel.org, rcu@vger.kernel.org
Subject: Re: [linus:master] [rcu/nocb]  7625926086:
 WARNING:at_kernel/rcu/tree_nocb.h:#rcu_nocb_try_bypass
Message-ID: <005a79ac-5080-40a5-a7bd-31c0d2e76414@paulmck-laptop>
Reply-To: paulmck@kernel.org
References: <202307111535.4823c255-oliver.sang@intel.com>
 <ZK1983kQCrN+zCqs@lothringen>
 <842683bc-5859-48be-8ca5-17a1e4bf3f39@paulmck-laptop>
 <ZK7b+vIJpOZhndbm@lothringen>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <ZK7b+vIJpOZhndbm@lothringen>
Precedence: bulk

On Wed, Jul 12, 2023 at 06:59:38PM +0200, Frederic Weisbecker wrote:
> On Tue, Jul 11, 2023 at 09:41:58AM -0700, Paul E. McKenney wrote:
> > On Tue, Jul 11, 2023 at 06:06:11PM +0200, Frederic Weisbecker wrote:
> > Heh!
> > 
> > The purpose was to see if this lock was ever contended.  I guess we now
> > have an answer, which is "Yes, but rarely."
> > 
> > It looks like the victim commit increased the size of the ->nocb_lock
> > critical section, just enough to make this happen sometimes.
> > 
> > Removing the WARN_ON_ONCE() seems appropriate, especially given that
> > this only happens when shrinking.
> 
> Ok, I'll check that.
> 
> > Should we add something that monitors that lock's contention?  It is
> > often the case that lock contention rises over time as new features and
> > optimizations are added.
> 
> I'm not sure. Should we keep the current ->nocb_lock_contended based engine
> to report contention somehow somewhere? Also does it behave better than our
> current spinlock slow path implementations?

The current spinlock slow path implementations try to be fair, and here
we want to favor the things dequeuing from ->cblist over those enqueuing
to it.  The idea is to slow down the enqueuing when things get busy.

But to your point, given the rarity of contention, is this a real problem
needing to be solved?  My queuing-theory studies and experiences argue
strenuously in favor of keeping it, but maybe there is a better way.

							Thanx, Paul