[v3,3/4] clk: mediatek: Add pcw_chg_shift control

Commit Message

Daniel Golle Dec. 6, 2023, 12:57 a.m. UTC

From: Sam Shih <[e-mail]>

Introduce pcw_chg_shfit control to optionally use that instead of the
hardcoded PCW_CHG_MASK macro.
This will needed for clocks on the MT7988 SoC.

Signed-off-by: Sam Shih <[e-mail]>
Signed-off-by: Daniel Golle <[e-mail]>
---
v3: use git --from ...
v2: no changes

 drivers/clk/mediatek/clk-pll.c | 5 ++++-
 drivers/clk/mediatek/clk-pll.h | 1 +
 2 files changed, 5 insertions(+), 1 deletion(-)

Comments

AngeloGioacchino Del Regno Dec. 6, 2023, 10:38 a.m. UTC | #1

Il 06/12/23 01:57, Daniel Golle ha scritto:
> From: Sam Shih <[e-mail]>
> 
> Introduce pcw_chg_shfit control to optionally use that instead of the
> hardcoded PCW_CHG_MASK macro.
> This will needed for clocks on the MT7988 SoC.
> 
> Signed-off-by: Sam Shih <[e-mail]>
> Signed-off-by: Daniel Golle <[e-mail]>
> ---
> v3: use git --from ...
> v2: no changes
> 
>   drivers/clk/mediatek/clk-pll.c | 5 ++++-
>   drivers/clk/mediatek/clk-pll.h | 1 +
>   2 files changed, 5 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/clk/mediatek/clk-pll.c b/drivers/clk/mediatek/clk-pll.c
> index 513ab6b1b3229..9f08bc5d2a8a2 100644
> --- a/drivers/clk/mediatek/clk-pll.c
> +++ b/drivers/clk/mediatek/clk-pll.c
> @@ -114,7 +114,10 @@ static void mtk_pll_set_rate_regs(struct mtk_clk_pll *pll, u32 pcw,
>   			pll->data->pcw_shift);
>   	val |= pcw << pll->data->pcw_shift;
>   	writel(val, pll->pcw_addr);
> -	chg = readl(pll->pcw_chg_addr) | PCW_CHG_MASK;
> +	if (pll->data->pcw_chg_shift)
> +		chg = readl(pll->pcw_chg_addr) | BIT(pll->data->pcw_chg_shift);
> +	else
> +		chg = readl(pll->pcw_chg_addr) | PCW_CHG_MASK;
>   	writel(chg, pll->pcw_chg_addr);
>   	if (pll->tuner_addr)
>   		writel(val + 1, pll->tuner_addr);
> diff --git a/drivers/clk/mediatek/clk-pll.h b/drivers/clk/mediatek/clk-pll.h
> index f17278ff15d78..d28d317e84377 100644
> --- a/drivers/clk/mediatek/clk-pll.h
> +++ b/drivers/clk/mediatek/clk-pll.h
> @@ -44,6 +44,7 @@ struct mtk_pll_data {
>   	u32 pcw_reg;
>   	int pcw_shift;
>   	u32 pcw_chg_reg;
> +	int pcw_chg_shift;
>   	const struct mtk_pll_div_table *div_table;
>   	const char *parent_name;
>   	u32 en_reg;

Hmm... no, this is not the best at all and can be improved.

Okay, so, the situation here is that one or some PLL(s) on MT7988 have a different
PCW_CHG_MASK as far as I understand.

Situation here is:
  - Each PLL must be registered to clk-pll
  - Each driver declaring a PLL does exactly so
    - There's a function to register the PLL

You definitely don't want to add a conditional in pll_set_rate(): even though
this is technically not a performance path on the current SoCs (and will probably
never be), it's simply useless to have this (very small) overhead there.

The solution is to:
  - Change that pcw_chg_shift to an unsigned short int type (or u8, your call):
    you don't need 32 bits for this number, as the expected range of this member
    is [0-31], and this can be expressed in just 4 bits (u8 is the smallest though)
  - Add that to function mtk_clk_register_pll_ops()
  - Change mtk_pll_set_rate_regs() to always do
    chg = readl(pll->pcw_chg_addr) | BIT(pll->data->pcw_chg_shift);

Cheers,
Angelo

Daniel Golle Dec. 9, 2023, 12:10 a.m. UTC | #2

Hi Angelo,

thank you for taking the time to review and for the helpful comments.

On Wed, Dec 06, 2023 at 11:38:36AM +0100, AngeloGioacchino Del Regno wrote:
> Il 06/12/23 01:57, Daniel Golle ha scritto:
> > From: Sam Shih <[e-mail]>
> > 
> > Introduce pcw_chg_shfit control to optionally use that instead of the
> > hardcoded PCW_CHG_MASK macro.
> > This will needed for clocks on the MT7988 SoC.
> > 
> > Signed-off-by: Sam Shih <[e-mail]>
> > Signed-off-by: Daniel Golle <[e-mail]>
> > ---
> > v3: use git --from ...
> > v2: no changes
> > 
> >   drivers/clk/mediatek/clk-pll.c | 5 ++++-
> >   drivers/clk/mediatek/clk-pll.h | 1 +
> >   2 files changed, 5 insertions(+), 1 deletion(-)
> > 
> > diff --git a/drivers/clk/mediatek/clk-pll.c b/drivers/clk/mediatek/clk-pll.c
> > index 513ab6b1b3229..9f08bc5d2a8a2 100644
> > --- a/drivers/clk/mediatek/clk-pll.c
> > +++ b/drivers/clk/mediatek/clk-pll.c
> > @@ -114,7 +114,10 @@ static void mtk_pll_set_rate_regs(struct mtk_clk_pll *pll, u32 pcw,
> >   			pll->data->pcw_shift);
> >   	val |= pcw << pll->data->pcw_shift;
> >   	writel(val, pll->pcw_addr);
> > -	chg = readl(pll->pcw_chg_addr) | PCW_CHG_MASK;
> > +	if (pll->data->pcw_chg_shift)
> > +		chg = readl(pll->pcw_chg_addr) | BIT(pll->data->pcw_chg_shift);
> > +	else
> > +		chg = readl(pll->pcw_chg_addr) | PCW_CHG_MASK;
> >   	writel(chg, pll->pcw_chg_addr);
> >   	if (pll->tuner_addr)
> >   		writel(val + 1, pll->tuner_addr);
> > diff --git a/drivers/clk/mediatek/clk-pll.h b/drivers/clk/mediatek/clk-pll.h
> > index f17278ff15d78..d28d317e84377 100644
> > --- a/drivers/clk/mediatek/clk-pll.h
> > +++ b/drivers/clk/mediatek/clk-pll.h
> > @@ -44,6 +44,7 @@ struct mtk_pll_data {
> >   	u32 pcw_reg;
> >   	int pcw_shift;
> >   	u32 pcw_chg_reg;
> > +	int pcw_chg_shift;
> >   	const struct mtk_pll_div_table *div_table;
> >   	const char *parent_name;
> >   	u32 en_reg;
> 
> Hmm... no, this is not the best at all and can be improved.
> 
> Okay, so, the situation here is that one or some PLL(s) on MT7988 have a different
> PCW_CHG_MASK as far as I understand.

Correct. *All* clocks of MT7988 have a different PCW_CHG_MASK, BIT(2)
instead of BIT(31).

> 
> Situation here is:
>  - Each PLL must be registered to clk-pll
>  - Each driver declaring a PLL does exactly so
>    - There's a function to register the PLL
> 
> You definitely don't want to add a conditional in pll_set_rate(): even though
> this is technically not a performance path on the current SoCs (and will probably
> never be), it's simply useless to have this (very small) overhead there.
> 
> The solution is to:
>  - Change that pcw_chg_shift to an unsigned short int type (or u8, your call):
>    you don't need 32 bits for this number, as the expected range of this member
>    is [0-31], and this can be expressed in just 4 bits (u8 is the smallest though)

Ack will use u8 instead, despite the struct not being packed, so I
wonder if it actually makes a difference.

>  - Add that to function mtk_clk_register_pll_ops()
>  - Change mtk_pll_set_rate_regs() to always do
>    chg = readl(pll->pcw_chg_addr) | BIT(pll->data->pcw_chg_shift);

As mtk_pll_data is a read-only member of the mtk_pll struct, we can't
set pcw_chg_shift to 31 in mtk_clk_register_pll_ops() in case it
is set to 0.
The only (much more intrusive change) would be to explicitely declare
.pcw_chg_shift = 31 in all current drivers setting .pcs_chg_reg != 0.
Should I do that instead?

Patch

diff --git a/drivers/clk/mediatek/clk-pll.c b/drivers/clk/mediatek/clk-pll.c
index 513ab6b1b3229..9f08bc5d2a8a2 100644
--- a/drivers/clk/mediatek/clk-pll.c
+++ b/drivers/clk/mediatek/clk-pll.c
@@ -114,7 +114,10 @@  static void mtk_pll_set_rate_regs(struct mtk_clk_pll *pll, u32 pcw,
 			pll->data->pcw_shift);
 	val |= pcw << pll->data->pcw_shift;
 	writel(val, pll->pcw_addr);
-	chg = readl(pll->pcw_chg_addr) | PCW_CHG_MASK;
+	if (pll->data->pcw_chg_shift)
+		chg = readl(pll->pcw_chg_addr) | BIT(pll->data->pcw_chg_shift);
+	else
+		chg = readl(pll->pcw_chg_addr) | PCW_CHG_MASK;
 	writel(chg, pll->pcw_chg_addr);
 	if (pll->tuner_addr)
 		writel(val + 1, pll->tuner_addr);
diff --git a/drivers/clk/mediatek/clk-pll.h b/drivers/clk/mediatek/clk-pll.h
index f17278ff15d78..d28d317e84377 100644
--- a/drivers/clk/mediatek/clk-pll.h
+++ b/drivers/clk/mediatek/clk-pll.h
@@ -44,6 +44,7 @@  struct mtk_pll_data {
 	u32 pcw_reg;
 	int pcw_shift;
 	u32 pcw_chg_reg;
+	int pcw_chg_shift;
 	const struct mtk_pll_div_table *div_table;
 	const char *parent_name;
 	u32 en_reg;