Re: [PATCH v1 06/16] clk: starfive: Add JH7100 clock generator driver

[Andy Shevchenko <andy.shevchenko@xxxxxxxxx>]

On Tue, Oct 12, 2021 at 11:08 PM Emil Renner Berthing <kernel@xxxxxxxx> wrote:
> On Tue, 12 Oct 2021 at 17:40, Andy Shevchenko <andy.shevchenko@xxxxxxxxx> wrote:
> > On Tue, Oct 12, 2021 at 4:42 PM Emil Renner Berthing <kernel@xxxxxxxx> wrote:

...

> > > +       value |= readl_relaxed(reg) & ~mask;
> >
> > value is not masked, is it okay?
> >
> > Usual pattern for this kind of operations is
> >
> > value = (current & ~mask) | (value & mask);
>
> This function is only ever called with constants, already masked
> values or the parent number from the clk framework, so it should be
> ok.

Up to you, but I think it's better to have a usual pattern.

> > > +       writel_relaxed(value, reg);

...

> > > +       rate = parent / div;
> > > +       if (rate < req->min_rate && div > 1) {
> > > +               div -= 1;
> > > +               rate = parent / div;
> > > +       }
> >
> > Seems like homegrown DIV_ROUND_UP() or so. Who will guarantee that
> > decreasing div by 1 will satisfy the conditional again?
>
> Maths unless I'm mistaken: div = DIV_ROUND_UP(parent, target), so in
> rational numbers
>   div - 1 < parent / target
> But the target is clamped by min_rate and max_rate, so
>   min_rate <= target < parent / (div - 1) = rate
>
> Sorry, re-using the rate varable for both the target and result is
> confusing. I'll fix that.

Also needs a comment, I believe.

...

> > > +#ifdef CONFIG_DEBUG_FS
> >
> > Perhaps __maybe_unused?
>
> I can definitely use __maybe_unused for the function declaration, but
> then I'll need a conditional every time clk_ops.debug_init needs to be
> initialized to either the function or NULL depending on
> CONFIG_DEBUG_FS below. Is that better?

Actually, why can't you always initialize the field? Shouldn't CLK
core take care about this conditional?

> > > +#else
> > > +#define jh7100_clk_debug_init NULL
> > > +#endif

...

> > > +       while (idx > 0)
> > > +               clk_hw_unregister(&priv->reg[--idx].hw);
> >
> > The
> >
> >        while (idx--)
> >                clk_hw_unregister(&priv->reg[idx].hw);
> >
> > is slightly better to read.
>
> It's not something I'll insist hard on, but I must admit I disagree.
> To me the above looks like cartoon characters running off a cliff and
> back. As a middle ground could we maybe do this?
>
>   while (idx)
>     clk_hw_unregister(&priv->reg[--idx].hw);

My point is exactly in having the common pattern for error paths, i.e.

  while (counter--)
    ...bla-bla-bla...

Your second approach is better, but I think that proposed by me is even better.

...

> > > +subsys_initcall(clk_starfive_jh7100_init);
> >
> > Any  explanation why subsys_initcall() is in use?
>
> TBH I just inherited that from Geert's first mock driver and never
> thought to question it. What would be a better alternative to try?

At least add a comment to explain the choice.

-- 
With Best Regards,
Andy Shevchenko