Re: [PATCH v2 06/16] clk: starfive: Add JH7100 clock generator driver

[Heiko Stübner]

Am Mittwoch, 27. Oktober 2021, 12:24:07 CEST schrieb Emil Renner Berthing:
> On Wed, 27 Oct 2021 at 02:54, Stephen Boyd <sboyd@xxxxxxxxxx> wrote:
> > Quoting Emil Renner Berthing (2021-10-26 15:35:36)
> > > On Tue, 26 Oct 2021 at 22:20, Stephen Boyd <sboyd@xxxxxxxxxx> wrote:
> > > > Quoting Emil Renner Berthing (2021-10-21 10:42:13)
> > > > > +};
> > > > > +
> > > > > +struct clk_starfive_jh7100_priv {
> > > > > +       /* protect registers against overlapping read-modify-write */
> > > > > +       spinlock_t rmw_lock;
> > > >
> > > > Does overlapping mean concurrent?
> > >
> > > Yes, sorry.
> > >
> > > > Do different clks share the same registers?
> > >
> > > No, each clock has their own register, but they use that register both
> > > to gate the clock and other configuration. The Locking chapter of
> > > Documentation/driver-api/clk.rst talks about the prepare lock and the
> > > enable lock and then says:
> > > "However, access to resources that are shared between operations of
> > > the two groups needs to be protected by the drivers. An example of
> > > such a resource would be a register that controls both the clock rate
> > > and the clock enable/disable state."
> >
> > Alright got it. Maybe say "protect clk enable and set rate from
> > happening at the same time".
> >
> > >
> > > > > +               return ERR_PTR(-EINVAL);
> > > > > +       }
> > > > > +
> > > > > +       if (idx >= JH7100_CLK_PLL0_OUT)
> > > > > +               return priv->pll[idx - JH7100_CLK_PLL0_OUT];
> > > > > +
> > > > > +       return &priv->reg[idx].hw;
> > > > > +}
> > > > > +
> > > > > +static int __init clk_starfive_jh7100_probe(struct platform_device *pdev)
> > > >
> > > > Drop __init as this can be called after kernel init is over.
> > >
> > > Oh interesting, I'd like to know when that can happen. The comment for
> > > the builtin_platform_driver macro says it's just a wrapper for
> >
> > I thought this was using module_platform_driver() macro?
> >
> > > device_initcall.
> > >
> > > Won't we then need to remove all the __initconst tags too since the
> > > probe function walks through jh7100_clk_data which eventually
> > > references all __initconst data?
> >
> > Yes. If it's builtin_platform_driver() it can't be a module/tristate
> > Kconfig, in which case all the init markings can stay.
> 
> Yes, it's already bool in the Kconfig file. After looking into this I
> think it's better to do like the rockchip drivers and use
> builtin_platform_driver_probe to make sure the probe function only
> called at kernel init time:
> 
> static struct platform_driver clk_starfive_jh7100_driver = {
>         .driver = {
>                 .name = "clk-starfive-jh7100",
>                 .of_match_table = clk_starfive_jh7100_match,
>                 .suppress_bind_attrs = true,
>         },
> };
> builtin_platform_driver_probe(clk_starfive_jh7100_driver,
> clk_starfive_jh7100_probe);
> 
> @Andy: is the supress_bind_attrs what you were asking about?
> 
> > > > > +
> > > > > +               clk->hw.init = &init;
> > > > > +               clk->idx = idx;
> > > > > +               clk->max = jh7100_clk_data[idx].max;
> > > > > +
> > > > > +               ret = clk_hw_register(priv->dev, &clk->hw);
> > > >
> > > > Why not use devm_clk_hw_register()?
> > >
> > > I probably could. Just for my understanding that's just to avoid the
> > > loop on error below, because as a builtin driver the device won't
> > > otherwise go away, right?
> >
> > Yes
> >
> > >
> > > > > +               if (ret)
> > > > > +                       goto err;
> > > > > +       }
> > > > > +
> > > > > +       ret = devm_of_clk_add_hw_provider(priv->dev, clk_starfive_jh7100_get, priv);
> > > > > +       if (ret)
> > > > > +               goto err;
> > > > > +
> > > > > +       return 0;
> > > > > +err:
> > > > > +       while (idx)
> > > > > +               clk_hw_unregister(&priv->reg[--idx].hw);
> > > > > +       return ret;
> > > > > +}
> > > > > +
> > > > > +static const struct of_device_id clk_starfive_jh7100_match[] = {
> > > > > +       { .compatible = "starfive,jh7100-clkgen" },
> > > > > +       { /* sentinel */ }
> > > > > +};
> > > >
> > > > Please add MODULE_DEVICE_TABLE()
> > >
> > > Will do!
> >
> > If it's never going to be a module then don't add any module_* things.
> 
> So does that just mean no MODULE_DEVICE_TABLE or should I also remove
> MODULE_DESCRIPTION, MODULE_AUTHOR and MODULE_LICENSE? I'm just double
> checking because the rockchip drivers seem to have MODULE_DESCRIPTION
> and MODULE_LICENSE lines.

reading this I realized that the current implementation on the Rockchip side
is faulty :-) .

For one, there recently was a change for 5.16 that moved this to use
module_platform_driver, while keeping the initdata, which suggest that
the change was actually not tested at all in that context.

And secondly the kconfig is wrong to actually use tristate and should instead
use bool.

Of course this is also influenced by the recent GKI discussion [0] ;-)

I'm going to revert the one change and also adapt the kconfig to use bool
again.


Heiko

[0] https://lwn.net/Articles/872209/