Re: [PATCH v5 3/3] pwm: Add support for Xilinx AXI Timer

On Mon, Aug 16, 2021 at 07:51:17PM -0400, Sean Anderson wrote:

On 8/14/21 4:47 PM, Uwe Kleine-König wrote:

quoted

Hello Sean,

sorry for having you let waiting so long. Now here some more feedback:

On Mon, Jul 19, 2021 at 06:13:22PM -0400, Sean Anderson wrote:

quoted

+static int xilinx_pwm_apply(struct pwm_chip *chip, struct pwm_device *unused,
+			    const struct pwm_state *state)
+{
+	bool enabled;
+	struct xilinx_timer_priv *priv = xilinx_pwm_chip_to_priv(chip);
+	u32 tlr0, tlr1, tcsr0, tcsr1;
+	u64 period_cycles, duty_cycles;
+	unsigned long rate;
+
+	if (state->polarity != PWM_POLARITY_NORMAL)
+		return -EINVAL;
+
+	/*
+	 * To be representable by TLR, cycles must be between 2 and
+	 * priv->max + 2. To enforce this we can reduce the duty
+	 * cycle, but we may not increase it.
+	 */
+	rate = clk_get_rate(priv->clk);
+	period_cycles = mul_u64_u32_div(state->period, rate, NSEC_PER_SEC);

cool, I didn't know mul_u64_u32_div.

I didn't either. Alas, many useful functions like these have no
documentation...

quoted

Hmm, we still have a problem here if

	state->period * rate > 1000000000 * U64_MAX.

Note that this can only occur with rate > 1GHz (and period = U64_MAX).
The highest fmax in the datasheet is 300 MHz (on a very expensive FPGA).

Maybe it is more prudent to do

	period = min(state->period, ULONG_MAX * NSEC_PER_SEC)

Together with a check for rate being <= 300 MHz to be safe that's fine.

I think a period of 136 years is adequate :) This comparison also has
the advantage of being against const values.

*nod*

quoted

quoted

+static void xilinx_pwm_get_state(struct pwm_chip *chip,
+				 struct pwm_device *unused,
+				 struct pwm_state *state)
+{
+	struct xilinx_timer_priv *priv = xilinx_pwm_chip_to_priv(chip);
+	u32 tlr0, tlr1, tcsr0, tcsr1;
+
+	regmap_read(priv->map, TLR0, &tlr0);
+	regmap_read(priv->map, TLR1, &tlr1);
+	regmap_read(priv->map, TCSR0, &tcsr0);
+	regmap_read(priv->map, TCSR1, &tcsr1);
+	state->period = xilinx_timer_get_period(priv, tlr0, tcsr0);

xilinx_timer_get_period rounds down, this is however wrong for
.get_state().

Why is this wrong? I thought get_state should return values which would
not be rounded if passed to apply_state.

Consider a PWM that yields a period of π * $regval ns when a certain
register is programmed with the value $regval.

Consider the HW is programmed with regval = 317. The exact period is
995.8848711879644. If now .get_state() rounds down and returns 995 ns and
you feed that value back into .apply the new regval (assuming round down
in .apply(), too) this yields regval = 316. If however .get_state()
rounds up and returns 996, putting this value back into .apply() you get
the desired 317.

Best regards
Uwe

-- 
Pengutronix e.K.                           | Uwe Kleine-König            |
Industrial Linux Solutions                 | https://www.pengutronix.de/ |