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, 16 Feb 2022 16:20:41 -0500
From:   Phil Auld <pauld@redhat.com>
To:     Aaron Tomlin <atomlin@redhat.com>
Cc:     frederic@kernel.org, tglx@linutronix.de, mingo@kernel.org,
        mingo@redhat.com, atomlin@atomlin.com,
        linux-kernel@vger.kernel.org, linux-mm@kvack.org
Subject: Re: [RFC PATCH] tick/sched: Ensure quiet_vmstat() is called when the
 idle tick was stopped too
Message-ID: <Yg1qqbBlHwkInC8O@lorien.usersys.redhat.com>
References: <20220203214339.1889971-1-atomlin@redhat.com>
 <YfxVpEO+UJTC+a9e@lorien.usersys.redhat.com>
 <20220216143412.dwxjlkq4w2zeweld@ava.usersys.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20220216143412.dwxjlkq4w2zeweld@ava.usersys.com>
Precedence: bulk

On Wed, Feb 16, 2022 at 02:34:12PM +0000 Aaron Tomlin wrote:
> On Thu 2022-02-03 17:22 -0500, Phil Auld wrote:
> > As I said earlier, I don't think you want to call quiet_vmstat()
> > unconditionally. And I don't think this will catch the cases you are
> > trying to fix. Once the tick is stopped tick_nohz_stop_tick should not be
> > getting called again until it's been restarted.
> 
> Phil,
> 
> Sorry about the delay. If I understand correctly, I see a scenario by which
> tick_nohz_stop_tick() can be called on transition/or exit from idle (e.g.
> default_idle_call()):

No worries. It's possible that I am misunderstanding the issue still... :)

> 
> 	1.	The idle/or scheduling-clock was previously
> 		stopped
> 
> 	2.	It is considered safe for the scheduling-clock
> 	        tick to remain "stopped"/or omitted; no need to
> 		reprogram and enable a periodic tick
> 		(e.g. no queued/or expired pending timer)
> 
> 	  ...
>             do_idle
> 	      cpuidle_idle_call
> 	      {
> 
> 		...
> 
> 	.--     default_idle_call
> 	|         arch_cpu_idle
> 	|         goto exit_idle
>         |
>         |       exit_idle:
> 	|         __current_set_polling()
>         |
> 	|     }
>         |     tick_nohz_idle_exit()
>         |     {
>         |
>         |       tick_stopped = ts->tick_stopped
>         |
>         |       if (tick_stopped)
>         |         tick_nohz_idle_update_tick(ts, now)
>         |           if (tick_nohz_full_cpu(smp_processor_id()))
>         |             __tick_nohz_full_update_tick(ts, now)
>         |             {
>         |               int cpu = smp_processor_id()
>         |
>         |               if (can_stop_full_tick(cpu, ts))
>         |                 tick_nohz_stop_sched_tick(ts, cpu)
>         |                   if (tick_nohz_next_event(ts, cpu))
>         '--                   tick_nohz_stop_tick(ts, cpu)
>                       }
>               }
> 
> If I understand correctly, __tick_nohz_full_update_tick() can return with
> no changes to the current tick (e.g. expire time == KTIME_MAX), no?
>

Yeah, I missed that path, so tick_nohz_stop_tick() can get called with the tick
already stopped.  My concern was about calling quiet_vmstat() if the tick was not
stopped as per the comment on that function. But looking more closely at
tick_nohz_stop_tick() it won't be doing that with your patch.

If this is fixing the issue you are seeing (I don't remember if you had a
reproducible case or not) then I think this could be a good way to do it.

It does seem to rely on a few things lining up right to get to the call to
tick_nohz_stop_tick(). 


Cheers,
Phil


> 
> Kind regards,
> 
> -- 
> Aaron Tomlin
> 

--