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: <20230826095836.1138608-1-liaochang1@huawei.com>
 <20230828072347.ly23mbptu3yw4zkv@vireshk-i7> <bcac1b3c-a2f7-7f18-4c2f-5cf27f4e89b5@huawei.com>
 <20230828085248.sz6aljr5aln7j435@vireshk-i7>
In-Reply-To: <20230828085248.sz6aljr5aln7j435@vireshk-i7>
From:   "Rafael J. Wysocki" <rafael@kernel.org>
Date:   Mon, 28 Aug 2023 10:58:30 +0200
Message-ID: <CAJZ5v0iGikZ=JSA5Nyx5Dc4QunSC5BObNO5yzQh44UYjrtRKYg@mail.gmail.com>
Subject: Re: [PATCH] cpufreq: Fix the race condition while updating the
 transition_task of policy
To:     Viresh Kumar <viresh.kumar@linaro.org>
Cc:     "Liao, Chang" <liaochang1@huawei.com>, rafael@kernel.org,
        srivatsa.bhat@linux.vnet.ibm.com, linux-pm@vger.kernel.org,
        linux-kernel@vger.kernel.org
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: 8BIT
Precedence: bulk

On Mon, Aug 28, 2023 at 10:52 AM Viresh Kumar <viresh.kumar@linaro.org> wrote:
>
> On 28-08-23, 16:29, Liao, Chang wrote:
> > Task B does not necessarily go to sleep when it calls wait_event(), it depends on
> > the condition to wait for evaluate false or not. So there is a small race window
> > where Task A already set 'transition_ongoing' to false and Task B can cross wait_event()
> > immediately.
> >
> > wait_event:
> > do {
> >       might_sleep();
> >       if (condition) // !transition_ongoing
> >               break;
> >       __wait_event();
> > };
> >
> > I hope I do not miss something important in the code above.
>
> > Yes, if the CPU uses weak memroy model, it is possible for the instructions to be reordered.
> > therefore, it is a good idea to insert an smb() between these two lines if there is race here.
>
> Maybe it would be better to do this instead ?
>
> diff --git a/drivers/cpufreq/cpufreq.c b/drivers/cpufreq/cpufreq.c
> index 6b52ebe5a890..f11b01b25e8d 100644
> --- a/drivers/cpufreq/cpufreq.c
> +++ b/drivers/cpufreq/cpufreq.c
> @@ -455,8 +455,10 @@ void cpufreq_freq_transition_end(struct cpufreq_policy *policy,
>                             policy->cur,
>                             policy->cpuinfo.max_freq);
>
> +       spin_lock(&policy->transition_lock);
>         policy->transition_ongoing = false;
>         policy->transition_task = NULL;
> +       spin_unlock(&policy->transition_lock);
>
>         wake_up(&policy->transition_wait);
>  }
>
> --

I was about to suggest the same thing.

wake_up() is a full memory barrier only if it actually wakes up a task
and if it doesn't do that, without the locking the other task may see
a state in which transition_ongoing is false already and
transition_task is still NULL regardless of the relative ordering of
the statements before the wake_up() call.