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: <20230721121534.44328-1-ppbuk5246@gmail.com> <05e98227-77f4-4918-8f8e-2170a158e350@paulmck-laptop>
 <CAM7-yPRc4Z0siaiWf+fK2GEfjPMq5UaY13F4o19rU6NNUS5fNg@mail.gmail.com>
 <faf07eef-0a51-49f3-be48-0433952171ad@paulmck-laptop> <CAM7-yPTrHN1xPXWjSUrJeTEOuy78DpmL8ytUY+a4ZOekiAbnZg@mail.gmail.com>
 <31070aab-8665-44c4-8950-0631a777ef44@paulmck-laptop> <CAM7-yPSn8ietAJ8NKb0-VHDQhkHs73u--KFmO3rpTPUvcFuMvA@mail.gmail.com>
In-Reply-To: <CAM7-yPSn8ietAJ8NKb0-VHDQhkHs73u--KFmO3rpTPUvcFuMvA@mail.gmail.com>
From:   Z qiang <qiang.zhang1211@gmail.com>
Date:   Mon, 24 Jul 2023 11:21:04 +0800
Message-ID: <CALm+0cWeZnMUk8Lj_nF3Htd14czGcT_Yt71nVwJdGhTECQOAWg@mail.gmail.com>
Subject: Re: [PATCH] rcu: remove unnecessary check cpu_no_qs.norm on rcu_report_qs_rdp
To:     Yun Levi <ppbuk5246@gmail.com>
Cc:     paulmck@kernel.org, frederic@kernel.org, quic_neeraju@quicinc.com,
        joel@joelfernandes.org, osh@joshtriplett.org, boqun.feng@gmail.com,
        rostedt@goodmis.org, mathieu.desnoyers@efficios.com,
        jiangshanlai@gmail.com, rcu@vger.kernel.org,
        linux-kernel@vger.kernel.org
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

>
> Thanks for replying to reply Paul :)
>
> > And try testing with CONFIG_RCU_STRICT_GRACE_PERIOD=y and CONFIG_PREEMPT=n.
> > Though there might be better Kconfig options to use.  Those two come
> > immediately to mind.
>
> I've tested with this option via rcu torture.
> and it doesn't report any problems.
> and after commit 6d60ea03ac2d3 ("rcu: Report QS for outermost
> PREEMPT=n rcu_read_unlock() for strict GPs")
> it always makes cpu_no_qs.b.norm false whenever it calls
> rcu_report_qs_rdp in rcu_read_unlock.
>
> > But one critical piece is that softirq handlers, including the RCU_SOFTIRQ
> > handler rcu_core_si(), can be invoked upon return from interrupts.
>
> I think in that case, no problem because if it reports qs already,
> rcu_check_quiescent_state wouldn't call rcu_report_qs_rdp again.
> And if RCU_SOFTIRQ is called as soon as RCU interrupt is finished,
> it seems the fastest one to notify qs to related tree.
>
> > Another critical piece is that if a CPU is idle during any part of a
> > grace period, the grace-period kthread can report a quiescent state on
> > its behalf.
>
> I think
>     1) If timer interrupt is still programed,
>           - when rcu_sched_clock_irq first reports qs, no problem
>           - If qs is reported via grace period thread first,
> note_gp_chagned in rcu interrupt
>             will recognize this situation by setting core_needs_qs as false,
>             it doesn't call rcu_report_qs_rdp thou cpu_no_qs.b.norm is true.
>
>      2) If the timer interrupt isn't programmed,
>           - rcu_gp_kthreaad reports its qs, no problem.
>
> So, I think there's no problem removing
>       "rdp->cpu_no_qs.b.norm" check in rcu_report_qs_rdp.
> or wrap this condition check as WARN_ON_ONCE.
>
> > Does that help?
> Many thanks always :)
>


Hi Levi

For built with CONFIG_RCU_STRICT_GRACE_PERIOD=y and CONFIG_PREEMPT=n kernels
Consider the following scenario:

__rcu_read_unlock()
   -> rcu_read_unlock_strict()
        ->rdp = this_cpu_ptr(&rcu_data);
        ->rdp->cpu_no_qs.b.norm = false;

                 by interrupt and return invoke rcu_core():
                 ->rcu_check_quiescent_state()
                      ->rdp = raw_cpu_ptr(&rcu_data);
                      -> rcu_check_quiescent_state(rdp);
                            ->note_gp_changes(rdp);
                                -> __note_gp_changes(rnp, rdp)
                                start new gp
                                ->rdp->cpu_no_qs.b.norm = true;

        ->rcu_report_qs_rdp(rdp);
           ->if (rdp->cpu_no_qs.b.norm || ...)


Thanks
Zqiang


>
> --------
> SIncerely,
> Levi.