Message-ID: <a121266a-110a-8c4e-cd71-4e35165fc789@gmail.com>
Date:   Mon, 22 May 2023 22:35:59 +0200
From:   Heiner Kallweit <hkallweit1@...il.com>
To:     Dmitry Rokosov <ddrokosov@...rdevices.ru>
Cc:     Martin Blumenstingl <martin.blumenstingl@...glemail.com>,
        neil.armstrong@...aro.org, jbrunet@...libre.com,
        mturquette@...libre.com, sboyd@...nel.org, robh+dt@...nel.org,
        krzysztof.kozlowski+dt@...aro.org, khilman@...libre.com,
        jian.hu@...ogic.com, kernel@...rdevices.ru, rockosov@...il.com,
        linux-amlogic@...ts.infradead.org, linux-clk@...r.kernel.org,
        devicetree@...r.kernel.org, linux-kernel@...r.kernel.org,
        linux-arm-kernel@...ts.infradead.org
Subject: Re: [PATCH v14 6/6] clk: meson: a1: add Amlogic A1 Peripherals clock
 controller driver

On 22.05.2023 15:44, Dmitry Rokosov wrote:
> Heiner,
> 
> On Fri, May 19, 2023 at 06:10:50PM +0200, Heiner Kallweit wrote:
>> On 18.05.2023 22:04, Martin Blumenstingl wrote:
>>> Hi Dmitry,
>>>
>>> On Wed, May 17, 2023 at 12:34 PM Dmitry Rokosov
>>> <ddrokosov@...rdevices.ru> wrote:
>>> [...]
>>>>>> Additionally, the CCF determines the best ancestor based on how close
>>>>>> its rate is to the given one, based on arithmetic calculations. However,
>>>>>> we have independent knowledge that a certain clock would be better, with
>>>>>> less jitter and fewer intermediaries, which will likely improve energy
>>>>>> efficiency. Sadly, the CCF cannot take this into account.
>>>>> I agree that the implementation in CCF is fairly simple. There's ways
>>>>> to trick it though: IIRC if there are multiple equally suitable clocks
>>>>> it picks the first one. For me all of this has worked so far which is
>>>>> what makes me curious in this case (not saying that anything is wrong
>>>>> with your approach).
>>>>>
>>>>> Do you have a (real world) example where the RTC clock should be
>>>>> preferred over another clock?
>>>>>
>>>>
>>>> Yes, a real-life example is the need for a 32Khz clock for an external
>>>> wifi chip. There is one option to provide this clock with high
>>>> precision, which is RTC + GENCLK.
>>>>
>>>>> I'm thinking about the following scenario.
>>>>> PWM parents:
>>>>> - XTAL: 24MHz
>>>>> - sys: not sure - let's say 166.67MHz
>>>>> - RTC: 32kHz
>>>>>
>>>>> Then after that there's a divider and a gate.
>>>>>
>>>>> Let's say the PWM controller needs a 1MHz clock: it can take that from
>>>>> XTAL or sys. Since XTAL is evenly divisible to 1MHz CCF will pick that
>>>>> and use the divider.
>>>>> But let's say the PWM controller needs a 32kHz clock: CCF would
>>>>> automatically pick the RTC clock.
>>>>> So is your implementation there to cover let's say 1kHz where
>>>>> mathematically 24MHz can be divided evenly to 1kHz (and thus should
>>>>> not result in any jitter) but RTC gives better precision in the real
>>>>> world (even though it's off by 24Hz)?
>>>>>
>>>>
>>>> I don't think so. The highest precision that RTC can provide is from a
>>>> 32KHz rate only. However, I believe that a 1kHz frequency can also be
>>>> achieved by using xtal 24MHz with a divider, which can provide high
>>>> precision as well.
>>> Thank you again for the great discussion on IRC today.
>>> Here's my short summary so I don't forget before you'll follow up on this.
>>>
>>> In general there's two known cases where the RTC clock needs to be used:
>>> a) When using the GENCLK output of the SoC to output the 32kHz RTC
>>> clock and connect that to an SDIO WiFi chip clock input (this seems
>>> useful in my understanding because the RTC clock provides high
>>> precision)
>>> b) When using the PWM controller to output a 32kHz clock signal. In
>>> this case my understanding is that using the RTC clock as input to the
>>> PWM controller results in the best possible signal
>>>
>>> The second case won't be supported with Heiner's patches [0] that use
>>> CCF (common clock framework) in the PWM controller driver.
>>> In this series the parent clock is calculated using:
>>>   freq = div64_u64(NSEC_PER_SEC * (u64)0xffff, period);
>>>
>>> A 32kHz clock means a PWM period of 30518ns. So with the above
>>
>> To be precise: 30517,578125ns
>> What means that the PWM framework can't say "I want 32768Hz",
>> but just "I want something being very close to 32768Hz".
>> So what you need is some simple heuristic to interpret the
>> PWM request -> "PWM requests 30518ns, but supposedly it wants
>> 32768Hz"
>>
>> NSEC_PER_SEC / 30518 = 32767 (rounded down from 32767,547)
>> clk_round_rate(channel->clk, 32767) would return 0 (I *think*),
>> because it tries to find the next lower clock.
>>
>> The SoC families I'm familiar with have fclkin2 as PWM parent.
>> That's 1 GHz in my case, what results in a frequency of 32.767,547Hz
>> for period = 30518n.
>> What you're saying is that newer generations don't have PWM parents
>>> 24MHz any longer?
> 
> No, of course not. For example, a fixed PLL (with all fclk_divX
> settings) has rates higher than 24MHz. However, we need to consider the
> 'heavy' background of such PWM.
> 
> However, we have a "lightweight" clkin (special rtc32k) with a rate of
> 32kHz that we could potentially use as an input to produce a 32kHz
> output on the PWM lines. I don't see any reason why we should not
> support such special cases.
> 

Two more things to consider:
1. When wanting a 32kHz (well, 32768Hz) output with a 50% duty cycle,
   then we need hi=0 and lo=0 with a 64kHz input clock.
   See point 2 for an explanation of why 0 and not 1.
   Means we couldn't use the RTC input clock. Did you consider this?
   Or do I miss something?
2. Seems the PWM block internally increments hi and lo, except the
   constant_en bit is set on newer PWM block versions.
   For bigger cnt values the impact is negligible, but for very small
   values it's something we have to consider.
   This was one additional motivation for me to choose an input
   frequency that creates big cnt values.

>>
>>
>>> calculation the PWM driver is asking for a clock rate of >=2GHz.
>>> We concluded that letting the common clock framework choose the best
>>> possible parent (meaning: removing CLK_SET_RATE_NO_REPARENT here) can
>>> be a way forward.
>>> But this means that the PWM controller driver must try to find the
>>> best possible parent somehow. The easiest way we came up with
>>> (pseudo-code):
>>>   freq = NSEC_PER_SEC / period;
>>>   fin_freq = clk_round_rate(channel->clk, freq);
>>>   if (fin_freq != freq) {
>>>     freq = div64_u64(NSEC_PER_SEC * (u64)0xffff, period);
>>>     fin_freq = clk_round_rate(channel->clk, freq);
>>>   }
>>>
>>> The idea is: for a requested 32kHz signal the PWM period is 30518ns.
>>> The updated logic would find that there's a matching clock input and
>>> use that directly. If not: use the original logic as suggested by
>>> Heiner.
>>>
>>>
>>> Best regards,
>>> Martin
>>>
>>>
>>> [0] https://lore.kernel.org/linux-amlogic/9faca2e6-b7a1-4748-7eb0-48f8064e323e@gmail.com/
>>
> 

Powered by blists - more mailing lists