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;
MIME-Version: 1.0
References: <20230704082615.7415-1-qiang.zhang1211@gmail.com>
 <77d16cbb-b28c-4b13-aecb-9ad787559557@paulmck-laptop> <CALm+0cW6DojcxSnzf2mMXhXP-MHJ0aEvSqyh1TbtvBC4bzfsOw@mail.gmail.com>
 <087e3da2-ffd9-4cf4-8893-73f4edc3cfba@paulmck-laptop>
In-Reply-To: <087e3da2-ffd9-4cf4-8893-73f4edc3cfba@paulmck-laptop>
From:   Joel Fernandes <joel@joelfernandes.org>
Date:   Fri, 7 Jul 2023 13:16:31 -0400
Message-ID: <CAEXW_YSfMSQ=z-pEO16_jwZSYt5N+zJHFaDOxDsPNoojm1v4Sg@mail.gmail.com>
Subject: Re: [PATCH] srcu: Make srcu_might_be_idle() take early return if
 rcu_gp_is_normal() return true
To:     paulmck@kernel.org
Cc:     Z qiang <qiang.zhang1211@gmail.com>, frederic@kernel.org,
        quic_neeraju@quicinc.com, rcu@vger.kernel.org,
        linux-kernel@vger.kernel.org
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
Precedence: bulk

On Fri, Jul 7, 2023 at 12:05=E2=80=AFPM Paul E. McKenney <paulmck@kernel.or=
g> wrote:
>
> On Fri, Jul 07, 2023 at 06:28:29PM +0800, Z qiang wrote:
> > >
> > > On Tue, Jul 04, 2023 at 04:26:15PM +0800, Zqiang wrote:
> > > > When invoke synchronize_srcu(), in the srcu_might_be_idle(), the cu=
rrent
> > > > CPU's sdp->lock will be acquired to check whether there are pending
> > > > callbacks in the sdp->srcu_cblist, if there are no pending callback=
s,
> > > > probabilistically probe global state to decide whether to convert t=
o
> > > > synchronize_srcu_expedited() call. however, for the rcupdate.rcu_no=
rmal=3D1
> > > > kernels and after the rcu_set_runtime_mode() is called, invoke the
> > > > rcu_gp_is_normal() is always return true, this mean that invoke the
> > > > synchronize_srcu_expedited() always fall back to synchronize_srcu()=
,
> > > > so there is no need to acquire sdp->lock to check sdp->srcu_cblist =
and
> > > > probe global state in srcu_might_be_idle().
> > > >
> > > > This commit therefore make srcu_might_be_idle() return immediately =
if the
> > > > rcu_gp_is_normal() return true.
> > > >
> > > > Signed-off-by: Zqiang <qiang.zhang1211@gmail.com>
> > > > ---
> > > >  kernel/rcu/srcutree.c | 2 ++
> > > >  1 file changed, 2 insertions(+)
> > > >
> > > > diff --git a/kernel/rcu/srcutree.c b/kernel/rcu/srcutree.c
> > > > index 20d7a238d675..aea49cb60a45 100644
> > > > --- a/kernel/rcu/srcutree.c
> > > > +++ b/kernel/rcu/srcutree.c
> > > > @@ -1172,6 +1172,8 @@ static bool srcu_might_be_idle(struct srcu_st=
ruct *ssp)
> > > >       unsigned long tlast;
> > > >
> > > >       check_init_srcu_struct(ssp);
> > > > +     if (rcu_gp_is_normal())
> > > > +             return false;
> > >
> > > Again, thank you for looking into SRCU!
> > >
> > > I am not at all enthusiastic about this one.  With this change, the n=
ame
> > > srcu_might_be_idle() is no longer accurate.  Yes, the name could chan=
ge,
> > > but any name would be longer and more confusing.
> > >
> > > So unless there is a measureable benefit to this one on a production
> > > workload, I cannot justify taking it.
> > >
> > > Is there a measureable benefit?
> >
> > Hi, Paul
> >
> > I only find that for Preempt-RT kernel,  the rcu_normal_after_boot is
> > set by default:
> > static int rcu_normal_after_boot =3D IS_ENABLED(CONFIG_PREEMPT_RT);
> > This affects only rcu but also srcu, this make the synchronize_srcu() a=
nd
> > synchronize_srcu_expedited() always fall back to __synchronize_srcu(ssp=
, true),
> > this means that call the srcu_might_be_idle() is meaningless.
>
> I do understand that the current setup favors default kernel builds at
> runtime by a few low-cost instructions, and that your change favors,
> as you say, kernels built for real-time, kernels built for certain types
> of HPC workloads, and all kernels during a small time during boot.
>
> My question is instead whether any of this makes a measureable difference
> at the system level.
>
> My guess is "no, not even close", but the way to convince me otherwise
> would be to actually run the workload and kernels on real hardware and
> provide measurements showing a statistically significant difference that
> the workload(s) in question care(s) about.
>
> So what can you show me?
>
> And srcu_might_be_idle() is not meaningless in that situation, just
> ignored completely.  And that is in fact the nature and purpose of the
> C-language || operator.  ;-)

I agree with Paul, without any evidence of improvement, optimizing an
obvious slow path is a NAK.

thanks,

 - Joel