Hello,

On Fri, Oct 11, 2024 at 01:27:33PM +0300, Aleksandr Shubin wrote:

It's a bit unfortunate that SUN20I_PWM_CLK_CFG is passed hwpwm/2 while
SUN20I_PWM_CTL gets a plain hwpwm. I suggest to at least name the
parameters differently.

+/*
+ * SUN20I_PWM_MAGIC is used to quickly compute the values of the clock dividers
+ * div_m (SUN20I_PWM_CLK_CFG_DIV_M) & prescale_k (SUN20I_PWM_CTL_PRESCAL_K)
+ * without using a loop. These dividers limit the # of cycles in a period
+ * to SUN20I_PWM_PCNTR_SIZE by applying a scaling factor of
+ * 1/(div_m * (prescale_k + 1)) to the clock source.
+ *
+ * SUN20I_PWM_MAGIC is derived by solving for div_m and prescale_k
+ * such that for a given requested period,
+ *
+ * i) div_m is minimized for any prescale_k ≤ SUN20I_PWM_CTL_PRESCAL_K_MAX,
+ * ii) prescale_k is minimized.
+ *
+ * The derivation proceeds as follows, with val = # of cycles for requested
+ * period:
+ *
+ * for a given value of div_m we want the smallest prescale_k such that
+ *
+ * (val >> div_m) // (prescale_k + 1) ≤ 65536 (SUN20I_PWM_PCNTR_SIZE)

                                                  ^
I'd add a = here ---------------------------------' to make it clear
that SUN20I_PWM_PCNTR_SIZE is a description for 65536 and this isn't a
multiplication.

+ *
+ * This is equivalent to:
+ *
+ * (val >> div_m) ≤ 65536 * (prescale_k + 1) + prescale_k
+ * ⟺ (val >> div_m) ≤ 65537 * prescale_k + 65536
+ * ⟺ (val >> div_m) - 65536 ≤ 65537 * prescale_k
+ * ⟺ ((val >> div_m) - 65536) / 65537 ≤ prescale_k
+ *
+ * As prescale_k is integer, this becomes
+ *
+ * ((val >> div_m) - 65536) // 65537 ≤ prescale_k
+ *
+ * And is minimized at
+ *
+ * ((val >> div_m) - 65536) // 65537
+ *
+ * Now we pick the smallest div_m that satifies prescale_k ≤ 255
+ * (i.e SUN20I_PWM_CTL_PRESCAL_K_MAX),
+ *
+ * ((val >> div_m) - 65536) // 65537 ≤ 255
+ * ⟺ (val >> div_m) - 65536 ≤ 255 * 65537 + 65536
+ * ⟺ val >> div_m ≤ 255 * 65537 + 2 * 65536
+ * ⟺ val >> div_m < (255 * 65537 + 2 * 65536 + 1)
+ * ⟺ div_m = fls((val) / (255 * 65537 + 2 * 65536 + 1))
+ *
+ * Suggested by Uwe Kleine-König
+ */
+#define SUN20I_PWM_MAGIC			(255 * 65537 + 2 * 65536 + 1)
+
+struct sun20i_pwm_chip {
+	struct clk *clk_bus, *clk_hosc, *clk_apb;
+	struct reset_control *rst;

clk_bus and rst are only used in probe and so can be a local variable there.

+	void __iomem *base;
+	struct mutex mutex; /* Protect PWM apply state */

Since commit 1cc2e1faafb3 ("pwm: Add more locking") that currently waits
in next for the merge window the callbacks for a single chip are
serialized, so (unless I miss something) this mutex can be dropped.

+};
+
+static inline struct sun20i_pwm_chip *to_sun20i_pwm_chip(struct pwm_chip *chip)
+{
+	return pwmchip_get_drvdata(chip);
+}
+
+static inline u32 sun20i_pwm_readl(struct sun20i_pwm_chip *chip,
+				   unsigned long offset)
+{
+	return readl(chip->base + offset);
+}
+
+static inline void sun20i_pwm_writel(struct sun20i_pwm_chip *chip,
+				     u32 val, unsigned long offset)
+{
+	writel(val, chip->base + offset);
+}
+
+static int sun20i_pwm_get_state(struct pwm_chip *chip,
+				struct pwm_device *pwm,
+				struct pwm_state *state)
+{
+	struct sun20i_pwm_chip *sun20i_chip = to_sun20i_pwm_chip(chip);
+	u16 ent_cycle, act_cycle, prescale_k;
+	u64 clk_rate, tmp;
+	u8 div_m;
+	u32 val;
+
+	mutex_lock(&sun20i_chip->mutex);
+
+	val = sun20i_pwm_readl(sun20i_chip, SUN20I_PWM_CLK_CFG(pwm->hwpwm / 2));
+	div_m = FIELD_GET(SUN20I_PWM_CLK_CFG_DIV_M, val);
+	if (div_m > SUN20I_PWM_CLK_DIV_M_MAX)
+		div_m = SUN20I_PWM_CLK_DIV_M_MAX;
+
+	if (FIELD_GET(SUN20I_PWM_CLK_CFG_SRC, val) == 0)

SUN20I_PWM_CLK_CFG_SRC is two bits wide. Do all values != 0 mean APB?

+		clk_rate = clk_get_rate(sun20i_chip->clk_hosc);
+	else
+		clk_rate = clk_get_rate(sun20i_chip->clk_apb);
+
+	val = sun20i_pwm_readl(sun20i_chip, SUN20I_PWM_CTL(pwm->hwpwm));
+	state->polarity = (SUN20I_PWM_CTL_ACT_STA & val) ?
+			   PWM_POLARITY_NORMAL : PWM_POLARITY_INVERSED;
+
+	prescale_k = FIELD_GET(SUN20I_PWM_CTL_PRESCAL_K, val) + 1;
+
+	val = sun20i_pwm_readl(sun20i_chip, SUN20I_PWM_ENABLE);
+	state->enabled = (SUN20I_PWM_ENABLE_EN(pwm->hwpwm) & val) ? true : false;
+
+	val = sun20i_pwm_readl(sun20i_chip, SUN20I_PWM_PERIOD(pwm->hwpwm));
+
+	mutex_unlock(&sun20i_chip->mutex);
+
+	act_cycle = FIELD_GET(SUN20I_PWM_PERIOD_ACT_CYCLE, val);
+	ent_cycle = FIELD_GET(SUN20I_PWM_PERIOD_ENTIRE_CYCLE, val);
+
+	/*
+	 * The duration of the active phase should not be longer
+	 * than the duration of the period
+	 */
+	if (act_cycle > ent_cycle)
+		act_cycle = ent_cycle;
+
+	/*
+	 * We have act_cycle <= ent_cycle <= 0xffff, prescale_k <= 0x100,
+	 * div_m <= 8. So the multiplication fits into an u64 without
+	 * overflow.
+	 */
+	tmp = ((u64)(act_cycle) * prescale_k << div_m) * NSEC_PER_SEC;
+	state->duty_cycle = DIV_ROUND_UP_ULL(tmp, clk_rate);
+	tmp = ((u64)(ent_cycle) * prescale_k << div_m) * NSEC_PER_SEC;
+	state->period = DIV_ROUND_UP_ULL(tmp, clk_rate);
+
+	return 0;
+}
+
+static int sun20i_pwm_apply(struct pwm_chip *chip, struct pwm_device *pwm,
+			    const struct pwm_state *state)
+{
+	struct sun20i_pwm_chip *sun20i_chip = to_sun20i_pwm_chip(chip);
+	u64 bus_rate, hosc_rate, val, ent_cycle, act_cycle;
+	u32 clk_gate, clk_cfg, pwm_en, ctl, reg_period;
+	u32 prescale_k, div_m;
+	bool use_bus_clk;
+
+	guard(mutex)(&sun20i_chip->mutex);
+
+	pwm_en = sun20i_pwm_readl(sun20i_chip, SUN20I_PWM_ENABLE);
+
+	if (state->enabled != pwm->state.enabled) {
+		clk_gate = sun20i_pwm_readl(sun20i_chip, SUN20I_PWM_CLK_GATE);
+
+		if (!state->enabled) {
+			clk_gate &= ~SUN20I_PWM_CLK_GATE_GATING(pwm->hwpwm);
+			pwm_en &= ~SUN20I_PWM_ENABLE_EN(pwm->hwpwm);
+			sun20i_pwm_writel(sun20i_chip, pwm_en, SUN20I_PWM_ENABLE);
+			sun20i_pwm_writel(sun20i_chip, clk_gate, SUN20I_PWM_CLK_GATE);
+
+			return 0;
+		}
+	}
+
+	ctl = sun20i_pwm_readl(sun20i_chip, SUN20I_PWM_CTL(pwm->hwpwm));
+	clk_cfg = sun20i_pwm_readl(sun20i_chip, SUN20I_PWM_CLK_CFG(pwm->hwpwm / 2));
+	hosc_rate = clk_get_rate(sun20i_chip->clk_hosc);
+	bus_rate = clk_get_rate(sun20i_chip->clk_apb);
+	if (pwm_en & SUN20I_PWM_ENABLE_EN(pwm->hwpwm ^ 1)) {
+		/* if the neighbor channel is enabled, check period only */
+		use_bus_clk = FIELD_GET(SUN20I_PWM_CLK_CFG_SRC, clk_cfg) != 0;
+		val = mul_u64_u64_div_u64(state->period,
+					  (use_bus_clk ? bus_rate : hosc_rate),
+					  NSEC_PER_SEC);
+
+		div_m = FIELD_GET(SUN20I_PWM_CLK_CFG_DIV_M, clk_cfg);
+	} else {
+		/*
+		 * Select the clock source based on the period,
+		 * since bus_rate > hosc_rate, which means bus_rate
+		 * can provide a higher frequency than hosc_rate.
+		 */
+		use_bus_clk = false;
+		val = mul_u64_u64_div_u64(state->period, hosc_rate, NSEC_PER_SEC);
+		if (val <= 1) {

Why is val == 1 already problematic?

+			use_bus_clk = true;
+			val = mul_u64_u64_div_u64(state->period, bus_rate, NSEC_PER_SEC);
+			if (val <= 1)
+				return -EINVAL;
+		}
+		div_m = fls(DIV_ROUND_DOWN_ULL(val, SUN20I_PWM_MAGIC));
+		if (div_m > SUN20I_PWM_CLK_DIV_M_MAX)
+			return -EINVAL;
+
+		/* set up the CLK_DIV_M and clock CLK_SRC */
+		clk_cfg = FIELD_PREP(SUN20I_PWM_CLK_CFG_DIV_M, div_m);
+		clk_cfg |= FIELD_PREP(SUN20I_PWM_CLK_CFG_SRC, use_bus_clk);
+
+		sun20i_pwm_writel(sun20i_chip, clk_cfg, SUN20I_PWM_CLK_CFG(pwm->hwpwm / 2));
+	}

Does writing SUN20I_PWM_CLK_CFG already influence the output? If so this
needs mentioning in the Limitations paragraph at the driver's top as the
output might glitch more than suggested there currently.

+	/* calculate prescale_k, PWM entire cycle */
+	ent_cycle = val >> div_m;
+	prescale_k = DIV_ROUND_DOWN_ULL(ent_cycle, 65537);

A #define for 65537?

+	if (prescale_k > SUN20I_PWM_CTL_PRESCAL_K_MAX)
+		prescale_k = SUN20I_PWM_CTL_PRESCAL_K_MAX;
+
+	do_div(ent_cycle, prescale_k + 1);
+
+	/* for N cycles, PPRx.PWM_ENTIRE_CYCLE = (N-1) */
+	reg_period = FIELD_PREP(SUN20I_PWM_PERIOD_ENTIRE_CYCLE, ent_cycle - 1);

Is ent_cycle known to be ≥ 1?

+	/* set duty cycle */
+	val = mul_u64_u64_div_u64(state->duty_cycle,
+				  (use_bus_clk ? bus_rate : hosc_rate),
+				  NSEC_PER_SEC);
+	act_cycle = val >> div_m;
+	do_div(act_cycle, prescale_k + 1);
+
+	/*
+	 * The formula of the output period and the duty-cycle for PWM are as follows.
+	 * T period = PWM0_PRESCALE_K / PWM01_CLK * (PPR0.PWM_ENTIRE_CYCLE + 1)
+	 * T high-level = PWM0_PRESCALE_K / PWM01_CLK * PPR0.PWM_ACT_CYCLE
+	 * Duty-cycle = T high-level / T period
+	 */
+	reg_period |= FIELD_PREP(SUN20I_PWM_PERIOD_ACT_CYCLE, act_cycle);
+	sun20i_pwm_writel(sun20i_chip, reg_period, SUN20I_PWM_PERIOD(pwm->hwpwm));
+
+	ctl = FIELD_PREP(SUN20I_PWM_CTL_PRESCAL_K, prescale_k);
+	if (state->polarity == PWM_POLARITY_NORMAL)
+		ctl |= SUN20I_PWM_CTL_ACT_STA;
+
+	sun20i_pwm_writel(sun20i_chip, ctl, SUN20I_PWM_CTL(pwm->hwpwm));
+
+	if (state->enabled != pwm->state.enabled && state->enabled) {
+		clk_gate &= ~SUN20I_PWM_CLK_GATE_BYPASS(pwm->hwpwm);
+		clk_gate |= SUN20I_PWM_CLK_GATE_GATING(pwm->hwpwm);
+		pwm_en |= SUN20I_PWM_ENABLE_EN(pwm->hwpwm);
+		sun20i_pwm_writel(sun20i_chip, pwm_en, SUN20I_PWM_ENABLE);
+		sun20i_pwm_writel(sun20i_chip, clk_gate, SUN20I_PWM_CLK_GATE);
+	}

There is an early return at the start of the function for
state->enabled != pwm->state.enabled && !state->enabled
. So just checking for

	if (state->enabled != pwm->state.enabled)

is enough.

+
+	return 0;
+}
+
+static const struct pwm_ops sun20i_pwm_ops = {
+	.apply = sun20i_pwm_apply,
+	.get_state = sun20i_pwm_get_state,
+};
+
+static const struct of_device_id sun20i_pwm_dt_ids[] = {
+	{ .compatible = "allwinner,sun20i-d1-pwm" },
+	{ },

No comma after the sentinel entry please.

+};
+MODULE_DEVICE_TABLE(of, sun20i_pwm_dt_ids);
+
+static void sun20i_pwm_reset_ctrl_release(void *data)
+{
+	struct reset_control *rst = data;
+
+	reset_control_assert(rst);
+}
+
+static int sun20i_pwm_probe(struct platform_device *pdev)
+{
+	struct pwm_chip *chip;
+	struct sun20i_pwm_chip *sun20i_chip;
+	int ret;
+
+	chip = devm_pwmchip_alloc(&pdev->dev, 8, sizeof(*sun20i_chip));
+	if (IS_ERR(chip))
+		return PTR_ERR(chip);
+	sun20i_chip = to_sun20i_pwm_chip(chip);
+
+	sun20i_chip->base = devm_platform_ioremap_resource(pdev, 0);
+	if (IS_ERR(sun20i_chip->base))
+		return PTR_ERR(sun20i_chip->base);
+
+	sun20i_chip->clk_bus = devm_clk_get_enabled(&pdev->dev, "bus");
+	if (IS_ERR(sun20i_chip->clk_bus))
+		return dev_err_probe(&pdev->dev, PTR_ERR(sun20i_chip->clk_bus),
+				     "failed to get bus clock\n");

I like error messages starting with a capital letter. Would you mind
converting accordingly?

+	sun20i_chip->clk_hosc = devm_clk_get_enabled(&pdev->dev, "hosc");
+	if (IS_ERR(sun20i_chip->clk_hosc))
+		return dev_err_probe(&pdev->dev, PTR_ERR(sun20i_chip->clk_hosc),
+				     "failed to get hosc clock\n");
+
+	sun20i_chip->clk_apb = devm_clk_get_enabled(&pdev->dev, "apb");
+	if (IS_ERR(sun20i_chip->clk_apb))
+		return dev_err_probe(&pdev->dev, PTR_ERR(sun20i_chip->clk_apb),
+				     "failed to get apb clock\n");
+
+	if (clk_get_rate(sun20i_chip->clk_apb) > clk_get_rate(sun20i_chip->clk_hosc))
+		dev_info(&pdev->dev, "apb clock must be greater than hosc clock");
+
+	sun20i_chip->rst = devm_reset_control_get_exclusive(&pdev->dev, NULL);
+	if (IS_ERR(sun20i_chip->rst))
+		return dev_err_probe(&pdev->dev, PTR_ERR(sun20i_chip->rst),
+				     "failed to get bus reset\n");
+
+	ret = of_property_read_u32(pdev->dev.of_node, "allwinner,pwm-channels",
+				   &chip->npwm);

error checking for ret?

+	if (chip->npwm > 16) {
+		dev_info(&pdev->dev, "limiting number of PWM lines from %u to 16",
+			 chip->npwm);
+		chip->npwm = 16;

Layer violation; drivers are not supposed to assign npwm. Also above you
only allocated 8. Better check allwinner,pwm-channels before calling
devm_pwmchip_alloc().

+	}
+
+	/* Deassert reset */
+	ret = reset_control_deassert(sun20i_chip->rst);
+	if (ret)
+		return dev_err_probe(&pdev->dev, ret, "failed to deassert reset\n");
+
+	ret = devm_add_action_or_reset(&pdev->dev, sun20i_pwm_reset_ctrl_release, sun20i_chip->rst);
+	if (ret)
+		return ret;

There is devm_reset_control_get_exclusive_asserted() scheduled to go
into v6.13-rc1. Please group the operations concerning rst together such
that it can be converted trivially to that function.

+	chip->ops = &sun20i_pwm_ops;
+
+	mutex_init(&sun20i_chip->mutex);
+
+	ret = devm_pwmchip_add(&pdev->dev, chip);
+	if (ret < 0)
+		return dev_err_probe(&pdev->dev, ret, "failed to add PWM chip\n");
+
+	return 0;
+}

Best regards
Uwe