Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > linux.kernel > #1431028 > unrolled thread

Re: [PATCH v2 14/15] clk: sunxi-ng: Add H3 clocks

Back to article view | Back to linux.kernel

This discussion starts older than the indexed window; earlier articles aren't shown. The article labeled Started by below is the oldest one visible, not the original post.

Contents

  Re: [PATCH v2 14/15] clk: sunxi-ng: Add H3 clocks Michael Turquette <mturquette@baylibre.com> - 2016-06-25 02:30 +0200
    Re: [PATCH v2 14/15] clk: sunxi-ng: Add H3 clocks Maxime Ripard <maxime.ripard@free-electrons.com> - 2016-06-26 14:40 +0200
      Re: [PATCH v2 14/15] clk: sunxi-ng: Add H3 clocks Maxime Ripard <maxime.ripard@free-electrons.com> - 2016-06-28 10:40 +0200

#1431028 — Re: [PATCH v2 14/15] clk: sunxi-ng: Add H3 clocks

Hi Maxime,

Nice series! Looks really great to me. :-)

Quoting Maxime Ripard (2016-06-07 13:41:53)
> +static SUNXI_CCU_NKMP_WITH_GATE_LOCK(pll_cpux_clk, "pll-cpux",
> +                                    "osc24M", 0x000,
> +                                    8, 5,      /* N */
> +                                    4, 2,      /* K */
> +                                    0, 2,      /* M */
> +                                    16, 2,     /* P */
> +                                    BIT(31),   /* gate */
> +                                    BIT(28),   /* lock */
> +                                    0);

I'm more of a fan of expanding the struct with designated initializers
versus macro use, but that's only personal preference.

> +static const char * const ahb2_parents[] = { "ahb1" , "pll-periph0" };
> +static struct ccu_mux ahb2_clk = {
> +       .mux            = {
> +               .shift  = 0,
> +               .width  = 1,
> +
> +               .fixed_prediv   = {
> +                       .index  = 1,
> +                       .div    = 2,
> +               },
> +       },
> +
> +       .common         = {
> +               .reg            = 0x05c,
> +               .features       = CCU_FEATURE_FIXED_PREDIV,
> +               .hw.init        = SUNXI_HW_INIT_PARENTS("ahb2",
> +                                                       ahb2_parents,

Note that it's possible to initialize the parent strings here if you
prefer:

	.hw.init = &(struct clk_init_data){
		   .parent_names = (const char *[]){ "ahb1",
		   				     "pll-periph0" };

Similar to the above, no big deal, just an observation.

> +static struct ccu_common *sun8i_h3_ccu_clks[] = {
> +       [CLK_PLL_CPUX]          = &pll_cpux_clk.common,
> +       [CLK_PLL_AUDIO_BASE]    = &pll_audio_base_clk.common,
> +       [CLK_PLL_AUDIO]         = &pll_audio_clk.common,

OK, it looks like you followed the qcom clk driver approach here, which
is a nice way to do things. However, as Stephen alluded to in his
response to the cover letter, the clk_hw_* api's are an even more
friendly interface for clock providers. For example, check out the gxbb
clk driver probe:

	static int gxbb_clkc_probe(struct platform_device *pdev)
	{
		void __iomem *clk_base;
		int ret, clkid, i;
		struct device *dev = &pdev->dev;
	
		/*  Generic clocks and PLLs */
		clk_base = of_iomap(dev->of_node, 0);
		if (!clk_base) {
			pr_err("%s: Unable to map clk base\n", __func__);
			return -ENXIO;
		}
	
		/* Populate base address for PLLs */
		for (i = 0; i < ARRAY_SIZE(gxbb_clk_plls); i++)
			gxbb_clk_plls[i]->base = clk_base;
	
		/* Populate base address for MPLLs */
		for (i = 0; i < ARRAY_SIZE(gxbb_clk_mplls); i++)
			gxbb_clk_mplls[i]->base = clk_base;

		...

		/*
		 * register all clks
		 */
		for (clkid = 0; clkid < NR_CLKS; clkid++) {
			ret = devm_clk_hw_register(dev, gxbb_hw_onecell_data.hws[clkid]);
			if (ret)
				goto iounmap;
		}

The nice thing about struct ccu_common is that you don't have to walk
the list of clocks for each separate clock type like the above probe
function does. I'm still thinking of the best way to solve this
generically. Maybe add a .base member struct clk_hw? I dunno, and I've
resisted the urge to add stuff to struct clk_hw in the past... But I
really want to minimize this .probe as much as possible, and I do not
want every clock provider driver to be forced to invent something like
struct ccu_common every time. Anyways, that is not a blocker for your
implementation to be merged, but Stephen's question in patch #4 got me
thinking about this again...

The real nice part is the call to devm_clk_hw_register. That uses the
new clk_hw_* apis and struct clk_hw_onecell_data, which is initialized
statically like so:

	static struct clk_hw_onecell_data gxbb_hw_onecell_data = {
		.hws = {
			[CLKID_SYS_PLL]	     = &gxbb_sys_pll.hw,
			[CLKID_CPUCLK]	      = &gxbb_cpu_clk.hw,
			...
		},
		.num = NR_CLKS,
	};

Unfortunately I believe it impossible to replace NR_CLKS with some
ARRAY_SIZE stuff because C. As Stephen mentioned, please use this method
instead.

> diff --git a/include/dt-bindings/clock/sun8i-h3.h b/include/dt-bindings/clock/sun8i-h3.h
> new file mode 100644
> index 000000000000..96eced56e7a2
> --- /dev/null
> +++ b/include/dt-bindings/clock/sun8i-h3.h
> @@ -0,0 +1,162 @@
> +#ifndef _DT_BINDINGS_CLK_SUN8I_H3_H_
> +#define _DT_BINDINGS_CLK_SUN8I_H3_H_
> +
> +#define CLK_PLL_CPUX           0
> +#define CLK_PLL_AUDIO_BASE     1
> +#define CLK_PLL_AUDIO          2
> +#define CLK_PLL_AUDIO_2X       3
> +#define CLK_PLL_AUDIO_4X       4

Are you sure you want to expose all of these clocks as part of the ABI?
I exposed the bare minimum clocks for the gxbb driver in the DT shared
header (we can always add more later) and kept the rest internal to the
kernel source.

A side benefit of this is that NR_CLKS can live inside the kernel and
not be part of the binding.

Otherwise this series looks really great. Thanks for tackling such a
huge task!

Regards,
Mike

[toc] | [next] | [standalone]

#1431486

[Multipart message — attachments visible in raw view] — view raw

Hi Mike,

On Fri, Jun 24, 2016 at 05:28:37PM -0700, Michael Turquette wrote:
> Hi Maxime,
> 
> Nice series! Looks really great to me. :-)

Great :)

> Quoting Maxime Ripard (2016-06-07 13:41:53)
> > +static SUNXI_CCU_NKMP_WITH_GATE_LOCK(pll_cpux_clk, "pll-cpux",
> > +                                    "osc24M", 0x000,
> > +                                    8, 5,      /* N */
> > +                                    4, 2,      /* K */
> > +                                    0, 2,      /* M */
> > +                                    16, 2,     /* P */
> > +                                    BIT(31),   /* gate */
> > +                                    BIT(28),   /* lock */
> > +                                    0);
> 
> I'm more of a fan of expanding the struct with designated initializers
> versus macro use, but that's only personal preference.

Chen-Yu was very much in favour of having everything defined by
macros, I didn't care that much either, so I guess he wins :)

> > +static const char * const ahb2_parents[] = { "ahb1" , "pll-periph0" };
> > +static struct ccu_mux ahb2_clk = {
> > +       .mux            = {
> > +               .shift  = 0,
> > +               .width  = 1,
> > +
> > +               .fixed_prediv   = {
> > +                       .index  = 1,
> > +                       .div    = 2,
> > +               },
> > +       },
> > +
> > +       .common         = {
> > +               .reg            = 0x05c,
> > +               .features       = CCU_FEATURE_FIXED_PREDIV,
> > +               .hw.init        = SUNXI_HW_INIT_PARENTS("ahb2",
> > +                                                       ahb2_parents,
> 
> Note that it's possible to initialize the parent strings here if you
> prefer:
> 
> 	.hw.init = &(struct clk_init_data){
> 		   .parent_names = (const char *[]){ "ahb1",
> 		   				     "pll-periph0" };
>
> Similar to the above, no big deal, just an observation.

A significant bunch of our clocks have the same parent list. Passing
the parent array allows to reuse the same arrays for those, so I'd
really like that over having VA_ARGS based macros (and that would put
the list of parents at the end of the declaration, which seems weird).
 
> > +static struct ccu_common *sun8i_h3_ccu_clks[] = {
> > +       [CLK_PLL_CPUX]          = &pll_cpux_clk.common,
> > +       [CLK_PLL_AUDIO_BASE]    = &pll_audio_base_clk.common,
> > +       [CLK_PLL_AUDIO]         = &pll_audio_clk.common,
> 
> OK, it looks like you followed the qcom clk driver approach here, which
> is a nice way to do things. However, as Stephen alluded to in his
> response to the cover letter, the clk_hw_* api's are an even more
> friendly interface for clock providers. For example, check out the gxbb
> clk driver probe:
> 
> 	static int gxbb_clkc_probe(struct platform_device *pdev)
> 	{
> 		void __iomem *clk_base;
> 		int ret, clkid, i;
> 		struct device *dev = &pdev->dev;
> 	
> 		/*  Generic clocks and PLLs */
> 		clk_base = of_iomap(dev->of_node, 0);
> 		if (!clk_base) {
> 			pr_err("%s: Unable to map clk base\n", __func__);
> 			return -ENXIO;
> 		}
> 	
> 		/* Populate base address for PLLs */
> 		for (i = 0; i < ARRAY_SIZE(gxbb_clk_plls); i++)
> 			gxbb_clk_plls[i]->base = clk_base;
> 	
> 		/* Populate base address for MPLLs */
> 		for (i = 0; i < ARRAY_SIZE(gxbb_clk_mplls); i++)
> 			gxbb_clk_mplls[i]->base = clk_base;
> 
> 		...
> 
> 		/*
> 		 * register all clks
> 		 */
> 		for (clkid = 0; clkid < NR_CLKS; clkid++) {
> 			ret = devm_clk_hw_register(dev, gxbb_hw_onecell_data.hws[clkid]);
> 			if (ret)
> 				goto iounmap;
> 		}

Ok, I'll move the fixed factor clocks out of the common list, and
initialize the clk_hw_onedata_cell structure to register it.

> The nice thing about struct ccu_common is that you don't have to walk
> the list of clocks for each separate clock type like the above probe
> function does. I'm still thinking of the best way to solve this
> generically. Maybe add a .base member struct clk_hw? I dunno, and I've
> resisted the urge to add stuff to struct clk_hw in the past... But I
> really want to minimize this .probe as much as possible, and I do not
> want every clock provider driver to be forced to invent something like
> struct ccu_common every time.

We'd need a few more things (in this case) at least: the register
offset and a private field to store our flags.

> Anyways, that is not a blocker for your implementation to be merged,
> but Stephen's question in patch #4 got me thinking about this
> again...
> 
> The real nice part is the call to devm_clk_hw_register. That uses the
> new clk_hw_* apis and struct clk_hw_onecell_data, which is initialized
> statically like so:
> 
> 	static struct clk_hw_onecell_data gxbb_hw_onecell_data = {
> 		.hws = {
> 			[CLKID_SYS_PLL]	     = &gxbb_sys_pll.hw,
> 			[CLKID_CPUCLK]	      = &gxbb_cpu_clk.hw,
> 			...
> 		},
> 		.num = NR_CLKS,
> 	};
> 
> Unfortunately I believe it impossible to replace NR_CLKS with some
> ARRAY_SIZE stuff because C. As Stephen mentioned, please use this method
> instead.

That's unfortunate :/

> > diff --git a/include/dt-bindings/clock/sun8i-h3.h b/include/dt-bindings/clock/sun8i-h3.h
> > new file mode 100644
> > index 000000000000..96eced56e7a2
> > --- /dev/null
> > +++ b/include/dt-bindings/clock/sun8i-h3.h
> > @@ -0,0 +1,162 @@
> > +#ifndef _DT_BINDINGS_CLK_SUN8I_H3_H_
> > +#define _DT_BINDINGS_CLK_SUN8I_H3_H_
> > +
> > +#define CLK_PLL_CPUX           0
> > +#define CLK_PLL_AUDIO_BASE     1
> > +#define CLK_PLL_AUDIO          2
> > +#define CLK_PLL_AUDIO_2X       3
> > +#define CLK_PLL_AUDIO_4X       4
> 
> Are you sure you want to expose all of these clocks as part of the ABI?
> I exposed the bare minimum clocks for the gxbb driver in the DT shared
> header (we can always add more later) and kept the rest internal to the
> kernel source.

I thought about it, but that would require a third array with
basically the same clocks:

  * the ccu_common array to patch to set the lock and base pointers,
  * the list of clocks to register
  * the clk_hw_onecell_data to deal with the dt binding.

That seems a bit overkill.

Thanks!
Maxime

-- 
Maxime Ripard, Free Electrons
Embedded Linux, Kernel and Android engineering
http://free-electrons.com

[toc] | [prev] | [next] | [standalone]

#1432636

[Multipart message — attachments visible in raw view] — view raw

On Mon, Jun 27, 2016 at 05:53:37PM -0700, Michael Turquette wrote:
> > > The nice thing about struct ccu_common is that you don't have to walk
> > > the list of clocks for each separate clock type like the above probe
> > > function does. I'm still thinking of the best way to solve this
> > > generically. Maybe add a .base member struct clk_hw? I dunno, and I've
> > > resisted the urge to add stuff to struct clk_hw in the past... But I
> > > really want to minimize this .probe as much as possible, and I do not
> > > want every clock provider driver to be forced to invent something like
> > > struct ccu_common every time.
> > 
> > We'd need a few more things (in this case) at least: the register
> > offset and a private field to store our flags.
> 
> A bit of mumbling to myself below:
> 
> Hmm, upon further reflection, walking the list of ccu clocks is rather
> identical to walking the list of each clock type, as I do in the gxbb
> driver, where ccu_common is the one and only clock type.
> 
> So that in itself is not a big deal and isn't a problem that needs
> solving.
> 
> What needs solving is a way to populate base addresses for each clock at
> runtime in a way that does not *force* you to invent something like
> ccu_common if you do not need it.
> 
> You would hit this issue if you broke your common gate or divider clocks
> out and did not wrap them in the ccu_common structure. I solved this by
> overloading the ->reg value of each of the common types as static data,
> and then did the following when registering them:
> 
> 	/* Populate base address for gates */
> 	for (i = 0; i < ARRAY_SIZE(gxbb_clk_gates); i++)
> 		gxbb_clk_gates[i]->reg = clk_base +
> 			(u64)gxbb_clk_gates[i]->reg;
> 
> Any thoughts on how to fix this for other common gate types that need
> their base addresses populated from an OF node at runtime?

One obvious way to work around it would be to allow to store a regmap
directly in the clk_hw structure, and then you'll only need to keep
the register offset in your clock type, which is fine (I think?).

> > > > diff --git a/include/dt-bindings/clock/sun8i-h3.h b/include/dt-bindings/clock/sun8i-h3.h
> > > > new file mode 100644
> > > > index 000000000000..96eced56e7a2
> > > > --- /dev/null
> > > > +++ b/include/dt-bindings/clock/sun8i-h3.h
> > > > @@ -0,0 +1,162 @@
> > > > +#ifndef _DT_BINDINGS_CLK_SUN8I_H3_H_
> > > > +#define _DT_BINDINGS_CLK_SUN8I_H3_H_
> > > > +
> > > > +#define CLK_PLL_CPUX           0
> > > > +#define CLK_PLL_AUDIO_BASE     1
> > > > +#define CLK_PLL_AUDIO          2
> > > > +#define CLK_PLL_AUDIO_2X       3
> > > > +#define CLK_PLL_AUDIO_4X       4
> > > 
> > > Are you sure you want to expose all of these clocks as part of the ABI?
> > > I exposed the bare minimum clocks for the gxbb driver in the DT shared
> > > header (we can always add more later) and kept the rest internal to the
> > > kernel source.
> > 
> > I thought about it, but that would require a third array with
> > basically the same clocks:
> > 
> >   * the ccu_common array to patch to set the lock and base pointers,
> >   * the list of clocks to register
> >   * the clk_hw_onecell_data to deal with the dt binding.
> 
> "the list of clocks to register" and "the clk_hw_onecell_data to deal
> with the dt binding" are the same array.
> 
> You only need two arrays:
> 
> 1) the ccu_common init data
> 2) the clk_hw_onecell_data array of clk_hw pointers that points to
> clk_hw statically defined in the ccu_common array

Ah, so you're still exposing them (anyone could be free to access of
the hidden clocks if it knows what value to use), but you're hiding
them (the ID is not public).

That could work, I'll change that.

Thanks!
Maxime

-- 
Maxime Ripard, Free Electrons
Embedded Linux, Kernel and Android engineering
http://free-electrons.com

[toc] | [prev] | [standalone]

Back to top | Article view | linux.kernel

csiph-web