On Thu, 2018-10-25 at 22:58 +0200, Martin Blumenstingl wrote:

Hi Jerome,

On Thu, Oct 25, 2018 at 2:54 PM Jerome Brunet [off-list ref] wrote:
[snip]

quoted

quoted

quoted

quoted

+static void clk_regmap_div_init(struct clk_hw *hw)
+{
+ struct clk_regmap *clk = to_clk_regmap(hw);
+ struct clk_regmap_div_data *div = clk_get_regmap_div_data(clk);
+ unsigned int val;
+ int ret;
+
+ ret = regmap_read(clk->map, div->offset, &val);
+ if (ret)
+         return;

+ val &= (clk_div_mask(div->width) << div->shift);
+ if (!val)
+         regmap_update_bits(clk->map, div->offset,
+                            clk_div_mask(div->width) << div->shift,
+                            clk_div_mask(div->width));

This is wrong for several reasons:
* You should hard code the initial value in the driver.
* If shift is not 0, I doubt this will give the expected result.

The value 0x00 of divider means nand clock off then read/write nand register is forbidden.

That is not entirely true, you can access the clock register or you'd be in a
chicken and egg situation.

quoted

Should we set the initial value in nand driver, or in sub emmc clk driver?

In the nand driver, which is the consumer of the clock. see my previous comments
about it.

an old version of this series had the code still in the NAND driver
(by writing to the registers directly instead of using the clk API).
this looks pretty much like a "sclk-div" to me (as I commented in v3
of this series: [0]):
- value 0 means disabled
- positive divider values
- (probably no duty control, but that's optional as far as I
understand sclk-div)
- uses max divider value when enabling the clock

if switching to sclk-div works then we can get rid of some duplicate code

It is possible:
There is a couple of things to note though:

* sclk does not 'uses max divider value when enabling the clock': Since this
divider can gate, it needs to save the divider value when disabling, since the
divider value is no longer stored in the register,
On init, this cached value is  saved as it is. If the divider is initially
disabled, we have to set the cached value to something that makes sense, in case
the clock is enabled without a prior call to clk_set_rate().

So in sclk, the clock setting is not changed nor hard coded in init, and this is
a very important difference.
 
* Even if sclk zero value means gated, it is still a zero based divider, while
eMMC/Nand divider is one based. It this controller was to sclk, then something
needs to be done for this.

* Since sclk caches a value in its data, and there can multiple instance of eMMC
/NAND clock controller, some care must be taken when registering the data.

Both the generic divider and sclk could work here ... it's up to you Jianxin.

Regards
Martin


[0] https://patchwork.kernel.org/patch/10607157/#22238243