Date:   Wed, 22 Feb 2023 15:48:57 +0100
From:   Frederic Weisbecker <frederic@kernel.org>
To:     Thomas Gleixner <tglx@linutronix.de>
Cc:     LKML <linux-kernel@vger.kernel.org>,
        Alexey Dobriyan <adobriyan@gmail.com>,
        Peter Zijlstra <peterz@infradead.org>,
        Wei Li <liwei391@huawei.com>,
        Mirsad Goran Todorovac <mirsad.todorovac@alu.unizg.hr>,
        Yu Liao <liaoyu15@huawei.com>, Hillf Danton <hdanton@sina.com>,
        Ingo Molnar <mingo@kernel.org>
Subject: Re: [PATCH 7/7] selftests/proc: Remove idle time monotonicity
 assertions
Message-ID: <Y/YrWZ6GnwLcWlwW@lothringen>
References: <20230220124129.519477-1-frederic@kernel.org>
 <20230220124129.519477-8-frederic@kernel.org>
 <87k00covg2.ffs@tglx>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <87k00covg2.ffs@tglx>
Precedence: bulk

On Mon, Feb 20, 2023 at 10:53:17PM +0100, Thomas Gleixner wrote:
> On Mon, Feb 20 2023 at 13:41, Frederic Weisbecker wrote:
> 
> > Due to broken iowait task counting design (cf: comments above
> > get_cpu_idle_time_us() and nr_iowait()), it is not possible to provide
> > the guarantee that /proc/stat or /proc/uptime display monotonic idle
> > time values.
> >
> > Remove the selftests that verify the related wrong assumption so that
> > testers and maintainers don't spend more time on that.
> >
> > Reported-by: Yu Liao <liaoyu15@huawei.com>
> > Reported-by: Thomas Gleixner <tglx@linutronix.de>
> 
> I did really not ask you to remove the selftests alltogether.
> 
> Those tests check uptime and idle time. I asked you to remove the idle
> time monotonicity assertion.
> 
> uptime is really supposed to be monotonically increasing. It's based on
> CLOCK_BOOTTIME. If that goes backwards then we surely have more trouble
> than /proc/uptime.

Right, I naively thought it wasn't worth checking CLOCK_BOOTTIME monotonicity
alone... Anyway, kept that in the new version.

> 
> But it would be a good thing to change the test in the following way:
> 
>        ut1 = parse("/proc/uptime");
>        bt1 = clock_gettime(CLOCK_BOOTTIME);
>        assert(ut1 <= bt1);
>        ut2 = parse("/proc/uptime");
>        bt2 = clock_gettime(CLOCK_BOOTTIME);
>        assert(ut2 <= bt2);
>        assert(ut1 <= ut2);
>        ....
> 
> Hmm?

Makes sense, added to the new version.

Thanks!

> 
> Thanks,
> 
>         tglx