Re: [PATCHv2 2/2] clk: meson: g12a: set cpub_clk flags to CLK_IS_CRITICAL

From: Jerome Brunet <jbrunet@baylibre.com>
To: Martin Blumenstingl <martin.blumenstingl@googlemail.com>,
	Kevin Hilman <khilman@baylibre.com>
Cc: Anand Moon <linux.amoon@gmail.com>,
	Rob Herring <robh+dt@kernel.org>,
	Mark Rutland <mark.rutland@arm.com>,
	Neil Armstrong <narmstrong@baylibre.com>,
	Michael Turquette <mturquette@baylibre.com>,
	Stephen Boyd <sboyd@kernel.org>,
	devicetree@vger.kernel.org, linux-arm-kernel@lists.infradead.org,
	linux-amlogic@lists.infradead.org, linux-kernel@vger.kernel.org,
	linux-clk@vger.kernel.org
Subject: Re: [PATCHv2 2/2] clk: meson: g12a: set cpub_clk flags to CLK_IS_CRITICAL
Date: Mon, 23 Mar 2020 10:03:55 +0100	[thread overview]
Message-ID: <1jr1xjcuis.fsf@starbuckisacylon.baylibre.com> (raw)
In-Reply-To: <CAFBinCB2WXZNRg4wdFD0RJ5k4hHqcfAOCHemvHzZE42-Mo5vzA@mail.gmail.com>

On Fri 20 Mar 2020 at 00:39, Martin Blumenstingl <martin.blumenstingl@googlemail.com> wrote:

> Hi Kevin,
>
> On Mon, Mar 2, 2020 at 6:01 PM Kevin Hilman <khilman@baylibre.com> wrote:
> [...]
>> > updating flags to CLK_IS_CRITICAL which help enable all the parent for
>> > cpub_clk.
>>
>> With current mainline, I've tested DVFS using CPUfreq on both clusters
>> on odroid-n2, and both clusters are booting, so I don't understand the
>> need for this patch.
> I *think* there is a race condition at kernel boot between cpufreq and
> disabling orphaned clocks
> I'm not sure I fully understand it though and I don't have any G12B
> board to verify it
>
> my understanding is that u-boot runs Linux off CPU0 which is clocked by cpub_clk
> this means we need to keep cpub_clk enabled as long as Linux wants the
> CPU0 processor to be enabled (on 32-bit ARM platforms that would be
> smp_operations.cpu_{kill,die})
> cpufreq does not call clk_prepare_enable on the CPU clocks so this
> means that the orphaned clock cleanup mechanism can disable it "at any
> time",

If nothing calls enable the cpu clock while it is managed by Linux
(cpufreq), there might something worth fixing. Adding CLK_IS_CRITICAL
will mask an issue that is still not explained.

"at any time": absolutely not.
Disabling unused clocks, is done only once, at during the late_init
stage.

If your clock gets disabled later on, it means it has been turned on and
off by another driver (possibly due to probe deferral)

> killing everything running on CPU0 and CPU1 (which are both
> clocked by cpub_clk)
>
> I have no explanation why this depends on booting from SD or eMMC.
>
> for the 32-bit SoCs we have CLK_IS_CRITICAL on the CPU clock as well
> since commit 0dad1ec65bc30a
> on G12A we have CLK_IS_CRITICAL on the sys_pll clocks, however my
> understanding is that cpub_clk could also be fed by one of the
> fixed_pll derived clocks (which have a gate as well, which may or may
> not be turned off by the orphaned clock cleanup - that is pure
> speculation from my side though).

Yes there are other critical clocks on Amlogic, mostly because the SCP Fw
driver does not claim the clocks it depends on. At least we know why
this flag is set and there should be a comment associated with it.

ATM, the issue reported by Anand (anyone else ?) is not explained. 

>
>
> Martin