Re: [PATCH] clk: imx: composite-8m: avoid glitches when set_rate would be a no-op

From: Peng Fan
Date: Thu Aug 03 2023 - 22:05:40 EST

On 8/2/2023 2:30 PM, Ahmad Fatoum wrote:
Hello Peng,

On 02.08.23 03:25, Peng Fan wrote:

On 8/2/2023 12:27 AM, Ahmad Fatoum wrote:
Reconfiguring the clock divider to the exact same value is observed
on an i.MX8MN to often cause a short clock pause, probably because
the divider restarts counting from the time the register is written.

This issue doesn't show up normally, because the clock framework will
take care to not call set_rate when the clock rate is the same.
However, when we configure an upstream clock (e.g. an audio_pll), the
common code will call set_rate with the newly calculated rate on all
children. As the new rate is different, we enter set_rate and compute
the same divider values, write them back and cause the glitch (e.g.
on a SAI's MCLK).

The CCM root has glitch-free mux. When upstream pll freq change,
the child set rate will also touch the mux bit, since div and mux
in one register, so the mux logic will also function.

Per design, it is glitch free, so I not understand well why glitch.

When you configure pll, the downstream sai clk should still not be enabled, right?

- sai5 is running normally and divides Audio PLL out by 16.
- audio_pll1 is increased by 32 Hz -> only kdiv changes, so no glitch
- imx8m_clk_composite_divider_set_rate(sai5) is called with
32 / 16 = 2 Hz more
- imx8m_clk_composite_divider_set_rate computes same divider as before
and writes register
- divider starts counting from zero, so we have a longer clock pause
than usual, e.g. 40ns -> 125ns, external MCLK consumer doesn't like that at all.

Thanks for detailed explaination, I would expect write this down in commit message.

So it's not a glitch in the transient high frequency sense, but rather a transient
low frequency period. I can reword the commit message to s/glitch/clock pause/
if you prefer.

clock pause would be better, since the hardware design could avoid real glitch.


And yes, if mux is switched, we will probably get the same clock pause, but
that is not a problem for me currently, because we don't switch parents except
at boot up. Afterwards, only PLL is tuned.



To avoid the glitch, we skip writing the same value back again.

Fixes: d3ff9728134e ("clk: imx: Add imx composite clock")
Signed-off-by: Ahmad Fatoum <a.fatoum@xxxxxxxxxxxxxx>
  drivers/clk/imx/clk-composite-8m.c | 12 +++++++-----
  1 file changed, 7 insertions(+), 5 deletions(-)

diff --git a/drivers/clk/imx/clk-composite-8m.c b/drivers/clk/imx/clk-composite-8m.c
index cbf0d7955a00..3e9a092e136c 100644
--- a/drivers/clk/imx/clk-composite-8m.c
+++ b/drivers/clk/imx/clk-composite-8m.c
@@ -97,7 +97,7 @@ static int imx8m_clk_composite_divider_set_rate(struct clk_hw *hw,
      int prediv_value;
      int div_value;
      int ret;
-    u32 val;
+    u32 orig, val;
        ret = imx8m_clk_composite_compute_dividers(rate, parent_rate,
                          &prediv_value, &div_value);
@@ -106,13 +106,15 @@ static int imx8m_clk_composite_divider_set_rate(struct clk_hw *hw,
        spin_lock_irqsave(divider->lock, flags);
  -    val = readl(divider->reg);
-    val &= ~((clk_div_mask(divider->width) << divider->shift) |
-            (clk_div_mask(PCG_DIV_WIDTH) << PCG_DIV_SHIFT));
+    orig = readl(divider->reg);
+    val = orig & ~((clk_div_mask(divider->width) << divider->shift) |
+               (clk_div_mask(PCG_DIV_WIDTH) << PCG_DIV_SHIFT));
        val |= (u32)(prediv_value  - 1) << divider->shift;
      val |= (u32)(div_value - 1) << PCG_DIV_SHIFT;
-    writel(val, divider->reg);
+    if (val != orig)
+        writel(val, divider->reg);
        spin_unlock_irqrestore(divider->lock, flags);