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;
From:   Kevin Hilman <khilman@baylibre.com>
To:     Christian Hewitt <christianshewitt@gmail.com>
Cc:     Rob Herring <robh+dt@kernel.org>,
        Mark Rutland <mark.rutland@arm.com>,
        Neil Armstrong <narmstrong@baylibre.com>,
        devicetree <devicetree@vger.kernel.org>,
        linux-arm-kernel@lists.infradead.org,
        linux-amlogic@lists.infradead.org, linux-kernel@vger.kernel.org
Subject: Re: [PATCH] arm64: dts: meson: remove CPU opps below 1GHz for G12B/SM1
In-Reply-To: <C70293C2-8CE8-42CE-911A-E0CC3DFB82FC@gmail.com>
References: <20220209135535.29547-1-christianshewitt@gmail.com>
 <7h7da3o6cg.fsf@baylibre.com>
 <C70293C2-8CE8-42CE-911A-E0CC3DFB82FC@gmail.com>
Date:   Fri, 11 Feb 2022 12:54:06 -0800
Message-ID: <7ha6exm8ep.fsf@baylibre.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Transfer-Encoding: quoted-printable
Precedence: bulk

Christian Hewitt <christianshewitt@gmail.com> writes:

>> On 10 Feb 2022, at 5:31 am, Kevin Hilman <khilman@baylibre.com> wrote:
>>=20
>> Christian Hewitt <christianshewitt@gmail.com> writes:
>>=20
>>> Amlogic G12B and SM1 devices experience CPU stalls and random board
>>> wedges when the system idles and CPU cores clock down to lower opp
>>> points. Recent vendor kernels include a change to remove 100-250MHz
>>> (with no explanation) [0] but other downstream sources also remove
>>> the 500/667MHz points (also with no explanation). Unless 100-667Mhz
>>> opps are removed or the CPU governor forced to performance, stalls
>>> are observed, so let's remove them an improve stability/uptime.
>>=20
>> Just curious: what CPUfreq governor do you use by default for the
>> LibreELEC kernel?
>
> LE uses ondemand. One of the original clues on the problem us that the
> issue isn=E2=80=99t seen in some of the retro-gaming forks on LE's codeba=
se
> which use the performance governor (and overclocks, etc.)

OK, thanks.  And does LE ever do full system suspend/resume?  Are things
stable for you across multiple suspend/resume cycles on G12B or SM1
devices?

I'm seeing hat with either powersave or performance, repeated
suspend/resume is stable, but with ondemand or schedultil it's not, even
with $SUBJECT patch applied.

If you have some time to test, seeing how long this loop[1] runs with
ondemand vs performance or powersave would be instructive.

Even more interesting...  if I set the governor to performance, but set
the suspend OPP to 1GHz[2] (which is what it would be for the powersave
governor), it is also unstable.  This suggests (to me) that any sort of
OPP change during the suspend/resume process is going to be unstable.
Now the challenge is to understand why so we can avoid it.

Thanks,

Kevin

[1]
while true;  do
    echo "=3D=3D=3D SUSPEND =3D=3D=3D"
    cat /sys/devices/system/cpu/cpufreq/policy0/scaling_governor
    cat /sys/devices/system/cpu/cpufreq/policy0/scaling_cur_freq
    cat /sys/devices/system/cpu/cpufreq/policy2/scaling_governor
    cat /sys/devices/system/cpu/cpufreq/policy2/scaling_cur_freq
    echo rtcwake -d rtc0 -m mem -s4
    echo "=3D=3D=3D RESUME =3D=3D=3D"
    sleep 4
done

[2]
diff --git a/arch/arm64/boot/dts/amlogic/meson-g12b-s922x.dtsi b/arch/arm64=
/boot/dts/amlogic/meson-g12b-s922x.dtsi
index 1e5d0ee5d541..37da8be85288 100644
--- a/arch/arm64/boot/dts/amlogic/meson-g12b-s922x.dtsi
+++ b/arch/arm64/boot/dts/amlogic/meson-g12b-s922x.dtsi
@@ -14,6 +14,7 @@ cpu_opp_table_0: opp-table-0 {
 		opp-100000000 {
 			opp-hz =3D /bits/ 64 <100000000>;
 			opp-microvolt =3D <731000>;
+ 		        opp-suspend;
 		};
=20
 		opp-250000000 {
@@ -79,6 +80,7 @@ cpub_opp_table_1: opp-table-1 {
 		opp-100000000 {
 			opp-hz =3D /bits/ 64 <100000000>;
 			opp-microvolt =3D <751000>;
+ 		        opp-suspend;
 		};
=20
 		opp-250000000 {