Message-ID: <CAGXv+5EyRROsh_=J1Fg4K+ZgfkERF4dh4R6WoGw9MnTBMNUCgQ@mail.gmail.com>
Date: Thu, 29 Feb 2024 18:34:03 +0800
From: Chen-Yu Tsai <wenst@...omium.org>
To: AngeloGioacchino Del Regno <angelogioacchino.delregno@...labora.com>
Cc: Pin-yen Lin <treapking@...omium.org>, Michael Turquette <mturquette@...libre.com>, 
	Stephen Boyd <sboyd@...nel.org>, Matthias Brugger <matthias.bgg@...il.com>, 
	Weiyi Lu <weiyi.lu@...iatek.com>, linux-kernel@...r.kernel.org, 
	linux-clk@...r.kernel.org, linux-mediatek@...ts.infradead.org, 
	linux-arm-kernel@...ts.infradead.org
Subject: Re: [PATCH v3 1/2] clk: mediatek: Introduce need_pm_runtime to mtk_clk_desc

On Thu, Feb 29, 2024 at 5:45 PM AngeloGioacchino Del Regno
<angelogioacchino.delregno@...labora.com> wrote:
>
> Il 29/02/24 08:17, Chen-Yu Tsai ha scritto:
> > On Mon, Feb 26, 2024 at 7:16 PM AngeloGioacchino Del Regno
> > <angelogioacchino.delregno@...labora.com> wrote:
> >>
> >> Il 23/02/24 05:27, Chen-Yu Tsai ha scritto:
> >>> On Mon, Jan 8, 2024 at 4:18 PM Pin-yen Lin <treapking@...omium.org> wrote:
> >>>>
> >>>> Introduce a new need_pm_runtime variable to mtk_clk_desc to indicate
> >>>> this clock controller needs runtime PM for its operations.
> >>>> Also do a runtime PM get on the clock controller during the
> >>>> probing stage to workaround a possible deadlock.
> >>>>
> >>>> Signed-off-by: Pin-yen Lin <treapking@...omium.org>
> >>>
> >>> Reviewed-by: Chen-Yu Tsai <wenst@...omium.org>
> >>>
> >>> The patch itself looks fine.
> >>>
> >>> Besides the MT8183 MFG clock issues, we do actually need this for the
> >>> MT8192 ADSP clock. Its power domain is not enabled by default.
> >>>
> >>
> >> ...but on MT8195 the ADSP clock works - because the ADSP node exists.
> >
> > That's an indirect dependency that should not be relied on. Say the clock
> > driver probed but the ADSP hasn't, and you try to read out the current
> > status. What would happen?
> >
> > - Read out works fine, because the power domain is default on, and hasn't
> >    been turned off by late cleanup
> > - Read out is bogus (but you can't tell)
> > - Read out hangs.
> >
> > The third is what happens on MT8192. There's still some issues on that
> > front, as even after I applied the ADSP power domain patches from MediaTek,
> > the readout was still hanging.
> >
>
> That MT8192 lockup story is getting crazy in my head... anyway, besides that,
> I get the point - I was somehow ignoring the fact that kernel modules do exist.
>
> Eh, sorry about that :-)
>
> >> This poses a question: should we make clock controllers depend on power domains,
> >> or should we keep everything powered off (hence clocks down - no power consumption)
> >> *unless* the user exists?
> >
> > That's a policy discussion separate from actual hardware dependencies.
> > *If* the clock controller needs the power domain to be active for the
> > registers to be accessed, the clock controller *must* have a direct
> > dependency on the power domain.
> >
>
> I admit I should've worded that better.
>
> "should we make clock controllers depend on power domains" was actually implying
> "IF those need one" :-)
>
> I really wonder if - at this point - it's simply a better idea to not restrict
> the call to devm_pm_runtime_enable/resume_and_get to `need_runtime_pm == true`.
>
> Do we really need to exclude that on other clock controllers that don't have
> any power domain dependency? Any side effect?
>
> Saying this because if we can avoid yet another per-SoC flag I'm really happy,
> as readability is also impacted and besides - if we ever find out that one of
> those need a power domain in the future, we'll need just one commit and just
> only in the devicetree, instead of enabling a flag in driver X as well as that,
> avoiding some (potentially unnecessary) noise... I guess.
>
> P.S.: I just noticed that the return value for the devm_pm_runtime_enable() call
>        is not being checked!
>
> .......
>
> In short....
>
> Chen-Yu, at this point, do you have any reason why we wouldn't be able and/or it
> wouldn't be a good idea to just avoid adding the `need_runtime_pm` flag (meaning
> that we perform pm_runtime calls for all clock drivers unconditionally)?
>
> If this is about longer boot time, I don't think that it's going to be more than
> a millisecond or two, so that should be completely ignorable.

I think it's just more of a "don't enable features you don't need" thing.
We already ran into a weird deadlock, which is why the devm_pm_runtime_enable()
call has that comment.

I don't think anyone has actually looked at it. As you said it shouldn't be
much, at least during boot time. It's one call per clock controller.

> Can you please do a test for that, or should I?

The earliest I can work on it would be some time next week. Does that work
for you?

ChenYu

> Cheers
> Angelo
>
> >> For the second one, this means that the *device* gets the power domain (adsp), and
> >> not the clock controller (which clocks are effectively useless if there's no user).
> >
> > No. See my previous paragraph.
> >
> > ChenYu
> >
> >> Angelo
> >>
> >>>> ---
> >>>>
> >>>> Changes in v3:
> >>>> - Update the commit message and the comments before runtime PM call
> >>>>
> >>>> Changes in v2:
> >>>> - Fix the order of error handling
> >>>> - Update the commit message and add a comment before the runtime PM call
> >>>>
> >>>>    drivers/clk/mediatek/clk-mtk.c | 19 +++++++++++++++++++
> >>>>    drivers/clk/mediatek/clk-mtk.h |  2 ++
> >>>>    2 files changed, 21 insertions(+)
> >>>>
> >>>> diff --git a/drivers/clk/mediatek/clk-mtk.c b/drivers/clk/mediatek/clk-mtk.c
> >>>> index 2e55368dc4d8..ba1d1c495bc2 100644
> >>>> --- a/drivers/clk/mediatek/clk-mtk.c
> >>>> +++ b/drivers/clk/mediatek/clk-mtk.c
> >>>> @@ -13,6 +13,7 @@
> >>>>    #include <linux/of.h>
> >>>>    #include <linux/of_address.h>
> >>>>    #include <linux/platform_device.h>
> >>>> +#include <linux/pm_runtime.h>
> >>>>    #include <linux/slab.h>
> >>>>
> >>>>    #include "clk-mtk.h"
> >>>> @@ -494,6 +495,18 @@ static int __mtk_clk_simple_probe(struct platform_device *pdev,
> >>>>                           return IS_ERR(base) ? PTR_ERR(base) : -ENOMEM;
> >>>>           }
> >>>>
> >>>> +
> >>>> +       if (mcd->need_runtime_pm) {
> >>>> +               devm_pm_runtime_enable(&pdev->dev);
> >>>> +               /*
> >>>> +                * Do a pm_runtime_resume_and_get() to workaround a possible
> >>>> +                * deadlock between clk_register() and the genpd framework.
> >>>> +                */
> >>>> +               r = pm_runtime_resume_and_get(&pdev->dev);
> >>>> +               if (r)
> >>>> +                       return r;
> >>>> +       }
> >>>> +
> >>>>           /* Calculate how many clk_hw_onecell_data entries to allocate */
> >>>>           num_clks = mcd->num_clks + mcd->num_composite_clks;
> >>>>           num_clks += mcd->num_fixed_clks + mcd->num_factor_clks;
> >>>> @@ -574,6 +587,9 @@ static int __mtk_clk_simple_probe(struct platform_device *pdev,
> >>>>                           goto unregister_clks;
> >>>>           }
> >>>>
> >>>> +       if (mcd->need_runtime_pm)
> >>>> +               pm_runtime_put(&pdev->dev);
> >>>> +
> >>>>           return r;
> >>>>
> >>>>    unregister_clks:
> >>>> @@ -604,6 +620,9 @@ static int __mtk_clk_simple_probe(struct platform_device *pdev,
> >>>>    free_base:
> >>>>           if (mcd->shared_io && base)
> >>>>                   iounmap(base);
> >>>> +
> >>>> +       if (mcd->need_runtime_pm)
> >>>> +               pm_runtime_put(&pdev->dev);
> >>>>           return r;
> >>>>    }
> >>>>
> >>>> diff --git a/drivers/clk/mediatek/clk-mtk.h b/drivers/clk/mediatek/clk-mtk.h
> >>>> index 22096501a60a..c17fe1c2d732 100644
> >>>> --- a/drivers/clk/mediatek/clk-mtk.h
> >>>> +++ b/drivers/clk/mediatek/clk-mtk.h
> >>>> @@ -237,6 +237,8 @@ struct mtk_clk_desc {
> >>>>
> >>>>           int (*clk_notifier_func)(struct device *dev, struct clk *clk);
> >>>>           unsigned int mfg_clk_idx;
> >>>> +
> >>>> +       bool need_runtime_pm;
> >>>>    };
> >>>>
> >>>>    int mtk_clk_pdev_probe(struct platform_device *pdev);
> >>>> --
> >>>> 2.43.0.472.g3155946c3a-goog
> >>>>
> >>
> >>
> >>
>
>
>

Powered by blists - more mailing lists