Re: [PATCH v3] clk: starfive: pll: Fix lower rate of CPUfreq by setting PLL0 rate to 1.5GHz

[Samuel Holland <samuel.holland@sifive.com>]

Hi Xingyu,

On 2024-04-03 2:44 AM, Xingyu Wu wrote:
> On 03/04/2024 15:24, Krzysztof Kozlowski wrote:
>>
>> On 03/04/2024 09:19, Xingyu Wu wrote:
>>> On 03/04/2024 0:18, Krzysztof Kozlowski wrote:
>>>>
>>>> On 02/04/2024 11:09, Xingyu Wu wrote:
>>>>> CPUfreq supports 4 cpu frequency loads on 375/500/750/1500MHz.
>>>>> But now PLL0 rate is 1GHz and the cpu frequency loads become
>>>>> 333/500/500/1000MHz in fact.
>>>>>
>>>>> So PLL0 rate should be default set to 1.5GHz. But setting the
>>>>> PLL0 rate need certain steps:
>>>>>
>>>>> 1. Change the parent of cpu_root clock to OSC clock.
>>>>> 2. Change the divider of cpu_core if PLL0 rate is higher than
>>>>>    1.25GHz before CPUfreq boot.
>>>>> 3. Change the parent of cpu_root clock back to PLL0 clock.
>>>>>
>>>>> Reviewed-by: Hal Feng <hal.feng@starfivetech.com>
>>>>> Fixes: e2c510d6d630 ("riscv: dts: starfive: Add cpu scaling for
>>>>> JH7110
>>>>> SoC")
>>>>> Signed-off-by: Xingyu Wu <xingyu.wu@starfivetech.com>
>>>>> ---
>>>>>
>>>>> Hi Stephen and Emil,
>>>>>
>>>>> This patch fixes the issue about lower rate of CPUfreq[1] by setting
>>>>> PLL0 rate to 1.5GHz.
>>>>>
>>>>> In order not to affect the cpu operation, setting the PLL0 rate need
>>>>> certain steps. The cpu_root's parent clock should be changed first.
>>>>> And the divider of the cpu_core clock should be set to 2 so they
>>>>> won't crash when setting 1.5GHz without voltage regulation. Due to
>>>>> PLL driver boot earlier than SYSCRG driver, cpu_core and cpu_root
>>>>> clocks are using by ioremap().
>>>>>
>>>>> [1]: https://github.com/starfive-tech/VisionFive2/issues/55
>>>>>
>>>>> Previous patch link:
>>>>> v2:
>>>>> https://lore.kernel.org/all/20230821152915.208366-1-xingyu.wu@starfi
>>>>> ve
>>>>> tech.com/
>>>>> v1:
>>>>> https://lore.kernel.org/all/20230811033631.160912-1-xingyu.wu@starfi
>>>>> ve
>>>>> tech.com/
>>>>>
>>>>> Thanks,
>>>>> Xingyu Wu
>>>>> ---
>>>>>  .../jh7110-starfive-visionfive-2.dtsi         |   5 +
>>>>>  .../clk/starfive/clk-starfive-jh7110-pll.c    | 102 ++++++++++++++++++
>>>>
>>>> Please do not mix DTS and driver code. That's not really portable.
>>>> DTS is being exported and used in other projects.
>>>
>>> OK, I will submit that in two patches.
>>>
>>>>
>>>> ...
>>>>
>>>>>
>>>>> @@ -458,6 +535,8 @@ static int jh7110_pll_probe(struct
>>>>> platform_device
>>>> *pdev)
>>>>>  	struct jh7110_pll_priv *priv;
>>>>>  	unsigned int idx;
>>>>>  	int ret;
>>>>> +	struct device_node *np;
>>>>> +	struct resource res;
>>>>>
>>>>>  	priv = devm_kzalloc(&pdev->dev, sizeof(*priv), GFP_KERNEL);
>>>>>  	if (!priv)
>>>>> @@ -489,6 +568,29 @@ static int jh7110_pll_probe(struct
>>>>> platform_device
>>>> *pdev)
>>>>>  			return ret;
>>>>>  	}
>>>>>
>>>>> +	priv->is_first_set = true;
>>>>> +	np = of_find_compatible_node(NULL, NULL,
>>>>> +"starfive,jh7110-syscrg");
>>>>
>>>> Your drivers should not do it. It's fragile, hides true link/dependency.
>>>> Please use phandles.
>>>>
>>>>
>>>>> +	if (!np) {
>>>>> +		ret = PTR_ERR(np);
>>>>> +		dev_err(priv->dev, "failed to get syscrg node\n");
>>>>> +		goto np_put;
>>>>> +	}
>>>>> +
>>>>> +	ret = of_address_to_resource(np, 0, &res);
>>>>> +	if (ret) {
>>>>> +		dev_err(priv->dev, "failed to get syscrg resource\n");
>>>>> +		goto np_put;
>>>>> +	}
>>>>> +
>>>>> +	priv->syscrg_base = ioremap(res.start, resource_size(&res));
>>>>> +	if (!priv->syscrg_base)
>>>>> +		ret = -ENOMEM;
>>>>
>>>> Why are you mapping other device's IO? How are you going to ensure
>>>> synced access to registers?
>>>
>>> Because setting PLL0 rate need specific steps and use the clocks of SYSCRG.
>>
>> That's not a reason to map other device's IO. That could be a reason for having
>> syscon or some other sort of relationship, like clock or reset.
>>
>>> But SYSCRG driver also need PLL clock to be clock source when adding
>>> clock providers. I tried to add SYSCRG clocks in 'clocks' property in
>>> DT and use
>>> clk_get() to get the clocks. But it could not run and crash. So I use
>>> ioremap() instead.
>>
>> So instead of properly model the relationship, you entangle the drivers even
>> more.
>>
>> Please come with a proper design for this. I have no clue about your hardware,
>> but that looks like you are asynchronously configuring the same hardware in two
>> different places.
>>
>> Sorry, that's poor code.
>>
>> Best regards,
>> Krzysztof
> 
> Hi Krzysztof,
> 
> If I use the old patch[1] like v2 and set the PLL0 default rate in the SYSCRG driver,
> will it be better?
> 
> [1]: https://lore.kernel.org/all/20230821152915.208366-1-xingyu.wu@starfivetech.com/

Both reparenting cpu_root and enforcing the maximum cpu_core frequency can be
accomplished with clk notifiers. See for example ccu_mux_notifier_register() in
drivers/clk/sunxi-ng/ccu_mux.c.

Regards,
Samuel