Re: [PATCH v6 2/4] pwm: opencores: Add PWM driver support

[William Qiu]

On 2023/11/2 19:30, Uwe Kleine-König wrote:
> Hello William,
> 
> On Wed, Nov 01, 2023 at 10:22:44AM +0800, William Qiu wrote:
>> 
>> 
>> On 2023/10/20 19:25, Uwe Kleine-König wrote:
>> >> +	void __iomem *base = pwm->data->get_ch_base ?
>> >> +			     pwm->data->get_ch_base(pwm->regs, dev->hwpwm) : pwm->regs;
>> >> +	u32 period_data, duty_data, ctrl_data;
>> >> +
>> >> +	period_data = readl(REG_OCPWM_LRC(base));
>> >> +	duty_data = readl(REG_OCPWM_HRC(base));
>> >> +	ctrl_data = readl(REG_OCPWM_CTRL(base));
>> >> +
>> >> +	state->period = DIV_ROUND_CLOSEST_ULL((u64)period_data * NSEC_PER_SEC, pwm->clk_rate);
>> >> +	state->duty_cycle = DIV_ROUND_CLOSEST_ULL((u64)duty_data * NSEC_PER_SEC, pwm->clk_rate);
>> > 
>> > Please test your driver with PWM_DEBUG enabled. The rounding is wrong
>> > here.
>> 
>> The conclusion after checking is: when the period or duty_cycle value set
>> by the user is not divisible (1000000000/49.5M), there will be an error.
>> This error is due to hardware accuracy. So why is rounding is wrong?
>> rockchip also has a similar implementation drivers/pwm/ pwm-rockchip.c
> 
> I fail to follow. Where is an error?
> 
> The general policy (for new drivers at least) is to implement the
> biggest period possible not bigger than the requested period. That means
> that .apply must round down and to make .apply ∘ .get_state idempotent
> .get_state must round up to match.
> 
> Assuming a clkrate of 49500000 Hz the actual period for REG_OCPWM_LRC =
> 400 is 8080.808ns and for REG_OCPWM_LRC = 401 is 8101.010.
> 
> So with REG_OCPWM_LRC = 401 .get_state should report state.period = 8102
> [ns] because if you call .apply with .period = 8101 [ns] you're supposed
> to use REG_OCPWM_LRC = 400.
> 
> Rounding using DIV_ROUND_CLOSEST doesn't give consistent behaviour in
> some cases. Consider a PWM that can implement the following periods (and
> none in between):
> 
> 	20.1 ns
> 	20.4 ns
> 	21.7 ns
> 
> With round-to-nearest a request to configure 21 ns will yield 20.4 ns.
> If you call .get_state there the driver will return 20 ns. However
> configuring 20 ns results in a period of 20.1 ns.
> 
> With rounding as requested above you get a consistent behaviour. After
> .apply_state(period=21) .get_state() returns period=21.
> 
> Best regards
> Uwe
> 
I see, then we'll use DIV_ROUND_DOWN_ULL for .apply() and DIV_ROUND_UP_ULL
for .get_state().
Thank you for your answer.

Best regards,
William