Re: [PATCH 4/5] clk: Add driver for Bitmain BM1880 SoC clock controller

[Stephen Boyd]

Quoting Manivannan Sadhasivam (2019-08-16 20:55:57)
> Hi Stephen,
> 
> On Wed, Aug 07, 2019 at 10:15:59PM -0700, Stephen Boyd wrote:
> > Quoting Manivannan Sadhasivam (2019-07-05 08:14:39)
> > > diff --git a/drivers/clk/Kconfig b/drivers/clk/Kconfig
> > > index fc1e0cf44995..ffc61ed85ade 100644
> > > --- a/drivers/clk/Kconfig
> > > +++ b/drivers/clk/Kconfig
> > > @@ -304,6 +304,12 @@ config COMMON_CLK_FIXED_MMIO
> > >         help
> > >           Support for Memory Mapped IO Fixed clocks
> > >  
> > > +config COMMON_CLK_BM1880
> > > +       bool "Clock driver for Bitmain BM1880 SoC"
> > > +       depends on ARCH_BITMAIN || COMPILE_TEST
> > > +       help
> > > +         This driver supports the clocks on Bitmain BM1880 SoC.
> > 
> > Can you add this config somewhere else besides the end? Preferably
> > close to alphabetically in this file.
> > 
> 
> Okay. I got confused by the fact that Makefile is sorted but not the
> Kconfig.

Ok. I'll make a reminder to sort the Kconfig after -rc1 next time.

> 
> > > +
> > >  source "drivers/clk/actions/Kconfig"
> > >  source "drivers/clk/analogbits/Kconfig"
> > >  source "drivers/clk/bcm/Kconfig"
> > > diff --git a/drivers/clk/clk-bm1880.c b/drivers/clk/clk-bm1880.c
> > > new file mode 100644
> > > index 000000000000..26cdb75bb936
> > > --- /dev/null
> > > +++ b/drivers/clk/clk-bm1880.c
[....]
> > > +
> > > +struct clk *bm1880_clk_register_pll(const struct bm1880_pll_clock *pll_clk,
> > > +                                   void __iomem *sys_base)
> > > +{
> > > +       struct bm1880_pll_hw_clock *pll_hw;
> > > +       struct clk_init_data init;
> > > +       struct clk_hw *hw;
> > > +       int err;
> > > +
> > > +       pll_hw = kzalloc(sizeof(*pll_hw), GFP_KERNEL);
> > > +       if (!pll_hw)
> > > +               return ERR_PTR(-ENOMEM);
> > > +
> > > +       init.name = pll_clk->name;
> > > +       init.ops = &bm1880_pll_ops;
> > > +       init.flags = pll_clk->flags;
> > > +       init.parent_names = &pll_clk->parent;
> > 
> > Can you use the new way of specifying parents instead of using strings
> > for everything?
> > 
> 
> Sure, will do it for clocks which doesn't use helper APIs.
> 
> > > +       init.num_parents = 1;
> > > +
> > > +       pll_hw->hw.init = &init;
> > > +       pll_hw->pll.reg = pll_clk->reg;
> > > +       pll_hw->base = sys_base;
> > > +
> > > +       hw = &pll_hw->hw;
> > > +       err = clk_hw_register(NULL, hw);
> > > +
> > > +       if (err) {
> > > +               kfree(pll_hw);
> > > +               return ERR_PTR(err);
> > > +       }
> > > +
> > > +       return hw->clk;
> > 
> > Can this return the clk_hw pointer instead?
> > 
> 
> What is the benefit? I see that only hw:init is going to be NULL in future.

Eventually we will remove ->clk from struct clk_hw and then this will
break. It also clearly makes this driver a clk provider driver and not a
clk consumer.

> So, I'll keep it as it is.

Please no!

> > > +               bm1880_clk_unregister_pll(data->clk_data.clks[clks[i].id]);
> > > +
> > > +       return PTR_ERR(clk);
> > > +}
> > > +
> > > +int bm1880_clk_register_mux(const struct bm1880_mux_clock *clks,
> > > +                           int num_clks, struct bm1880_clock_data *data)
> > > +{
> > > +       struct clk *clk;
> > > +       void __iomem *sys_base = data->sys_base;
> > > +       int i;
> > > +
> > > +       for (i = 0; i < num_clks; i++) {
> > > +               clk = clk_register_mux(NULL, clks[i].name,
> > 
> > Can you use the clk_hw based APIs for generic type clks?
> > 
> 
> IMO using helper APIs greatly reduce code size and makes the driver
> look more clean. So I prefer to use the helpers wherever applicable.
> When you plan to deprecate those, I'll switch over to plain clk_hw APIs.

We have clk_hw_register_mux(). Please use it. The clk based registration
APIs are deprecated.

> > > +       kfree(clk_data);
> > > +}
> > > +
> > > +CLK_OF_DECLARE_DRIVER(bm1880_clk, "bitmain,bm1880-clk", bm1880_clk_init);
> > 
> > Is there a reason why it can't be a platform driver?
> > 
> 
> Hmm, I looked into the majority of drivers which live under `driver/clk/`.
> Most of them are using CLK_OF_DECLARE_DRIVER, so I thought that only drivers
> which have a separate directory are preferred by the maintainers to use
> platform driver way.
> 
> Anyway, I can switch over to platform driver and that's what I prefer.
> 

Yes please use a platform driver unless it doesn't work for some reason.
Even then, use a platform driver and CLK_OF_DECLARE_DRIVER() in
conjunction to register the early clks from the OF_DECLARED section and
then adopt the rest to the proper device driver later on. This way we
gain the benefits of driver core.