Re: [PATCH] clk: bcm: Add driver for Northstar ILP clock

[Mark Rutland]

On Fri, Jul 29, 2016 at 02:58:32PM +0200, RafaÅ MiÅecki wrote:
> From: RafaÅ MiÅecki <rafal@xxxxxxxxxx>
> 
> This clock is present on cheaper Northstar devices like BCM53573 or
> BCM47189 using Corex-A7. This driver uses PMU (Power Management Unit)
> to calculate clock rate and allows using it in a generic (clk_*) way.
> 
> Signed-off-by: RafaÅ MiÅecki <rafal@xxxxxxxxxx>
> ---
>  .../devicetree/bindings/clock/brcm,ns-ilp.txt      |  28 ++++
>  drivers/clk/bcm/Makefile                           |   1 +
>  drivers/clk/bcm/clk-ns-ilp.c                       | 146 +++++++++++++++++++++
>  3 files changed, 175 insertions(+)
>  create mode 100644 Documentation/devicetree/bindings/clock/brcm,ns-ilp.txt
>  create mode 100644 drivers/clk/bcm/clk-ns-ilp.c
> 
> diff --git a/Documentation/devicetree/bindings/clock/brcm,ns-ilp.txt b/Documentation/devicetree/bindings/clock/brcm,ns-ilp.txt
> new file mode 100644
> index 0000000..c4df38e
> --- /dev/null
> +++ b/Documentation/devicetree/bindings/clock/brcm,ns-ilp.txt
> @@ -0,0 +1,28 @@
> +Broadcom Northstar ILP clock
> +============================
> +
> +This binding uses the common clock binding:
> +    Documentation/devicetree/bindings/clock/clock-bindings.txt
> +
> +This binding is used for ILP clock on Broadcom Northstar devices using
> +Corex-A7 CPU. ILP clock depends on ALP one and has to be calculated on
> +runtime.
> +
> +Required properties:
> +- compatible: "brcm,ns-ilp"
> +- reg: iomem address range of PMU (Power Management Unit)
> +- reg-names: "pmu", the only needed & supported reg right now

>From the commit message and binding description, it sounds like there
should be a binding for the PMU, and that should cover the clocks
required/exported by the PMU.

> +- clocks: should reference an ALP clock
> +- clock-names: "alp", the only needed & supported clock right now
> +- #clock-cells: should be <0>

How many clocks does the PMU output, including the ILP clock?

> +static unsigned long ns_ilp_recalc_rate(struct clk_hw *hw,
> +					unsigned long parent_rate)
> +{
> +	struct ns_ilp *ilp = container_of(hw, struct ns_ilp, hw);
> +	void __iomem *pmu = ilp->pmu;
> +	u32 last_val, cur_val;
> +	u32 sum = 0, num = 0, loop_num = 0;
> +	u32 avg;
> +	int err;
> +
> +	err = clk_prepare_enable(ilp->alp_clk);
> +	if (err)
> +		return 0;
> +
> +	/* Enable */
> +	writel(XTAL_CTL_EN, pmu + PMU_XTAL_FREQ_RATIO);

What exactly are we enabling here?

> +
> +	/* Read initial value */
> +	last_val = readl(pmu + PMU_XTAL_FREQ_RATIO) & XTAL_ALP_PER_4ILP;
> +
> +	/* Try getting 20 different values for calculating average */

Please describe *why* this is necessary (i.e. why we need to calculate
an average).

> +	while (num < 20) {
> +		cur_val = readl(pmu + PMU_XTAL_FREQ_RATIO) & XTAL_ALP_PER_4ILP;
> +
> +		if (cur_val != last_val) {
> +			/* Got different value, use it */
> +			sum += cur_val;
> +			num++;
> +			loop_num = 0;
> +			last_val = cur_val;
> +		} else if (++loop_num > 5000) {
> +			/* Same value over and over, give up */
> +			sum += cur_val;
> +			num++;
> +			break;
> +		}
> +	}
> +
> +	/* Disable */
> +	writel(0x0, pmu + PMU_XTAL_FREQ_RATIO);
> +
> +	avg = sum / num;
> +
> +	return clk_get_rate(ilp->alp_clk) * 4 / avg;
> +}

Shouldn't the clk_prepare_enable be balanced?

> +	index = of_property_match_string(np, "reg-names", "pmu");
> +	if (index < 0) {
> +		err = index;
> +		goto err_free_ilp;
> +	}
> +	err = of_address_to_resource(np, index, &res);
> +	if (err) {
> +		err = index;
> +		goto err_free_ilp;
> +	}
> +	ilp->pmu = ioremap_nocache(res.start, resource_size(&res));

Use ioremap() for mapping devices.

It is not correct to use ioremap_nocache().

We should probably add an of_get_address_resource_by_name() or something
like that to avoid drivers having to open-code this kind of thing.

> +	if (IS_ERR(ilp->pmu)) {
> +		err = PTR_ERR(ilp->pmu);
> +		goto err_free_ilp;
> +	}
> +
> +	ilp->alp_clk = of_clk_get_by_name(np, "alp-clk");

The binding document said "alp", not "alp-clk".

We already know it's a clock, so the "-clk" suffix is redundant.

> +	if (IS_ERR(ilp->alp_clk)) {
> +		err = PTR_ERR(ilp->alp_clk);
> +		goto err_unmap_pmu;
> +	}
> +
> +	init.name = np->name;
> +	init.ops = &ns_ilp_clk_ops;
> +	init.flags = CLK_IS_ROOT;

That's not true; the ALP clock is a parent...

I thought CLK_IS_ROOT was disappearing, regardless.

Thanks,
Mark,