Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
From:   Thomas Gleixner <tglx@linutronix.de>
To:     Will Deacon <will@kernel.org>,
        Mika =?utf-8?Q?Penttil=C3=A4?= <mika.penttila@nextfour.com>
Cc:     linux-arm-kernel@lists.infradead.org, linux-kernel@vger.kernel.org,
        Frederic Weisbecker <fweisbec@gmail.com>,
        Marc Zyngier <maz@kernel.org>,
        Lorenzo Colitti <lorenzo@google.com>,
        John Stultz <john.stultz@linaro.org>,
        Stephen Boyd <sboyd@kernel.org>, kernel-team@android.com
Subject: Re: [PATCH 3/5] tick/broadcast: Prefer per-cpu oneshot wakeup timers to broadcast
In-Reply-To: <20210521112503.GA11850@willie-the-truck>
References: <20210520184705.10845-1-will@kernel.org> <20210520184705.10845-4-will@kernel.org> <a269c869-b966-75d5-5fe1-6ed6921c1b83@nextfour.com> <20210521112503.GA11850@willie-the-truck>
Date:   Fri, 21 May 2021 15:49:39 +0200
Message-ID: <87tumw9oy4.ffs@nanos.tec.linutronix.de>
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Transfer-Encoding: quoted-printable
Precedence: bulk

On Fri, May 21 2021 at 12:25, Will Deacon wrote:
> On Fri, May 21, 2021 at 05:25:41AM +0300, Mika Penttil=C3=A4 wrote:
>> On 20.5.2021 21.47, Will Deacon wrote:
>> >   /*
>> >    * Conditionally install/replace broadcast device
>> >    */
>> > -void tick_install_broadcast_device(struct clock_event_device *dev)
>> > +void tick_install_broadcast_device(struct clock_event_device *dev, in=
t cpu)
>> >   {
>> >   	struct clock_event_device *cur =3D tick_broadcast_device.evtdev;
>> > +	if (tick_set_oneshot_wakeup_device(dev, cpu))
>> > +		return;
>> > +
>> >   	if (!tick_check_broadcast_device(cur, dev))
>> >   		return;
>>=20
>> Does this disable hpet registering as a global broadcast device on x86 ?=
 I
>> think it starts with cpumask =3D cpu0 so it qualifies for a percpu wakeup
>> timer.
>
> Well spotted, I think you're probably right. I'll try to reproduce on my
> laptop to confirm, but I hadn't noticed the tricks played with the cpumask
> on x86.
>
> I'll probably need to rework things so that we install the broadcast timer
> first, but prefer global devices.

HPET has cpumask(0) but does not have CLOCK_EVT_FEAT_PERCPU set. The
feature flag is a clear indicator for per cpu.

Thanks,

        tglx