Re: [PATCH RESEND 3/7] clk: spacemit: add reset controller support

[Haylen Chu <heylenay@4d2.org>]

On Fri, Mar 21, 2025 at 10:18:26AM -0500, Alex Elder wrote:
> Define ccu_reset_data as a structure that contains the constant
> register offset and bitmasks used to assert and deassert a reset
> control on a SpacemiT K1 CCU. Define ccu_reset_controller_data as
> a structure that contains the address of an array of those structures
> and a count of the number of elements in the array.
> 
> Add a pointer to a ccu_reset_controller_data structure to the
> k1_ccu_data structure.  Reset support is optional for SpacemiT CCUs;
> the new pointer field will be null for CCUs without any resets.
> 
> Finally, define a new ccu_reset_controller structure, which (for
> a CCU with resets) contains a pointer to the constant reset data,
> the regmap to be used for the controller, and an embedded a reset
> controller structure.
> 
> Each reset control is asserted or deasserted by updating bits in
> a register.  The bits used are defined by an assert mask and a
> deassert mask.  In some cases, one (non-zero) mask asserts reset
> and a different (non-zero) mask deasserts it.  Otherwise one mask
> is nonzero, and the other is zero.  Either way, the bits in
> both masks are cleared, then either the assert mask or the deassert
> mask is set in a register to affect the state of a reset control.
> 
> Signed-off-by: Alex Elder <elder@riscstar.com>
> ---
>  drivers/clk/spacemit/ccu-k1.c | 93 +++++++++++++++++++++++++++++++++++
>  1 file changed, 93 insertions(+)
> 
> diff --git a/drivers/clk/spacemit/ccu-k1.c b/drivers/clk/spacemit/ccu-k1.c
> index f7367271396a0..6d879411c6c05 100644
> --- a/drivers/clk/spacemit/ccu-k1.c
> +++ b/drivers/clk/spacemit/ccu-k1.c

...

> +static int
> +k1_rst_update(struct reset_controller_dev *rcdev, unsigned long id, bool assert)
> +{
> +	struct ccu_reset_controller *controller = rcdev_to_controller(rcdev);
> +	struct regmap *regmap = controller->regmap;
> +	const struct ccu_reset_data *data;
> +	u32 val;
> +	int ret;
> +
> +	data = &controller->data->data[id];
> +
> +	ret = regmap_read(regmap, data->offset, &val);
> +	if (ret)
> +		return ret;
> +
> +	val &= ~(data->assert_mask | data->deassert_mask);
> +	val |= assert ? data->assert_mask : data->deassert_mask;
> +
> +	return regmap_write(regmap, data->offset, val);
> +}

I don't think it's safe to write the regmap based on a value read
earlier without the regmap's inner lock held: it's totally fine for the
clock part to issue an update of the register at the same time. Without
knowledge on it, reset code may rollback the clock bits written by clock
code earlier to the original value. That's why I keep using ccu_update()
everywhere and dropped ccu_write().

Thanks,
Haylen Chu