Re: [PATCH v1 1/2] pinctrl: mediatek: support access registers without race-condition

From: Light Hsieh
Date: Fri Aug 21 2020 - 00:53:25 EST

On Wed, 2020-08-19 at 16:11 -0700, Sean Wang wrote:
> Hi Light,
>
> On Tue, Aug 18, 2020 at 1:36 AM <light.hsieh@xxxxxxxxxxxx> wrote:
> >
> > From: Light Hsieh <light.hsieh@xxxxxxxxxxxx>
> >
> > Some MediaTek SOC provide more control registers other than value register.
>
> s/MT6765/Some MediaTek SoC/
>
> > Generanll, a value register need read-modify-write is at offset 0xXXXXXXXX0.
>
> s/Generally/Generanll/
>
> > A corresponding SET register is at offset 0xXXXXXXX4. Write 1s' to some bits
> > of SET register will set same bits in value register.
> > A corresponding CLR register is at offset 0xXXXXXXX8. Write 1s' to some bits
> > of CLR register will clear same bits in value register.
> > For GPIO mode selection, MWR register is provided at offset 0xXXXXXXXC.
> > With MWR, the MSBit of GPIO mode selection field is for modification-enable,
> > not for GPIO mode selection, and the remaining LSBits are for mode
> > selection.
> > Take mode selection field with 4-bits as example, to select mode 0~7 via
> > MWR register, 8~15 (instead of 0~7) shall be written to corresponding mode
> > selection field.
> > When using SET/CLR/MWR registers, read-modify-write of value register is not
> > necessary. This can prevent from race condition when multiple bus masters
> > concurrently read-modify-write the same value register for setting different
> > fields of the same value register.
> >
> > Signed-off-by: Light Hsieh <light.hsieh@xxxxxxxxxxxx>
> > ---
> > drivers/pinctrl/mediatek/pinctrl-mtk-common-v2.c | 69 ++++++++++++++++++++++--
> > drivers/pinctrl/mediatek/pinctrl-mtk-common-v2.h | 2 +
> > 2 files changed, 67 insertions(+), 4 deletions(-)
> >
> > diff --git a/drivers/pinctrl/mediatek/pinctrl-mtk-common-v2.c b/drivers/pinctrl/mediatek/pinctrl-mtk-common-v2.c
> > index b77b18f..51f0b53 100644
> > --- a/drivers/pinctrl/mediatek/pinctrl-mtk-common-v2.c
> > +++ b/drivers/pinctrl/mediatek/pinctrl-mtk-common-v2.c
> > @@ -18,6 +18,29 @@
> > #include "mtk-eint.h"
> > #include "pinctrl-mtk-common-v2.h"
> >
> > +/* Some MediaTek SOC provide more control registers other than value register.
>
> s/MT6765/Some MediaTek SoC/

Not only MT6765 provides such control registers.
Actually, many (but not all) MediaTek SoC support.
Other MediaTek SoC can enable such control according to its HW support.

>
> > + * Generanll, a value register need read-modify-write is at offset 0xXXXXXXXX0.
>
> s/Generally/Generanll/
>
> > + * A corresponding SET register is at offset 0xXXXXXXX4. Write 1s' to some bits
> > + * of SET register will set same bits in value register.
> > + * A corresponding CLR register is at offset 0xXXXXXXX8. Write 1s' to some bits
> > + * of CLR register will clear same bits in value register.
> > + * For GPIO mode selection, MWR register is provided at offset 0xXXXXXXXC.
> > + * With MWR, the MSBit of GPIO mode selection field is for modification-enable,
> > + * not for GPIO mode selection, and the remaining LSBits are for mode
> > + * selection.
> > + * Take mode selection field with 4-bits as example, to select mode 0~7 via
> > + * MWR register, 8~15 (instead of 0~7) shall be written to corresponding mode
> > + * selection field.
> > + * When using SET/CLR/MWR registers, read-modify-write of value register is not
> > + * necessary. This can prevent from race condition when multiple bus masters
> > + * concurrently read-modify-write the same value register for setting different
> > + * fields of the same value register.
> > + */
> > +
> > +#define SET_OFFSET 0x4
> > +#define CLR_OFFSET 0x8
>
> can set/clr offset work for mode register?

Yes. However, use set/clr to change mode require 2 register access when
target mode is not all 0's or all 1's.
The mwr HW support is not available on mode register.

>
> > +#define MWR_OFFSET 0xC
> > +
> > /**
> > * struct mtk_drive_desc - the structure that holds the information
> > * of the driving current
> > @@ -64,6 +87,38 @@ void mtk_rmw(struct mtk_pinctrl *pctl, u8 i, u32 reg, u32 mask, u32 set)
> > mtk_w32(pctl, i, reg, val);
> > }
> >
> > +
> > +static void mtk_hw_set_value_race_free(struct mtk_pinctrl *pctl,
> > + struct mtk_pin_field *pf, u32 value)
>
> s/mtk_hw_set_value_race_free/mtk_hw_w1sc/ to explictly indicate
> write-one ethier set or clear operation supported by hw
>
> > +{
> > + unsigned int set, clr;
> > +
> > + set = value & pf->mask;
> > + clr = (~set) & pf->mask;
> > +
> > + if (set)
> > + mtk_w32(pctl, pf->index, pf->offset + SET_OFFSET,
> > + set << pf->bitpos);
> > + if (clr)
> > + mtk_w32(pctl, pf->index, pf->offset + CLR_OFFSET,
> > + clr << pf->bitpos);
> > +}
> > +
> > +static void mtk_hw_set_mode_race_free(struct mtk_pinctrl *pctl,
> > + struct mtk_pin_field *pf, u32 value)
>
> s/mtk_hw_set_mode_race_free/mtk_hw_mwr/
>
> > +{
> > + unsigned int value_new;
> > +
> > + /* MSB of mask is modification-enable bit, set this bit */
> > + value_new = (1 << (pctl->soc->mwr_field_width - 1)) | value;
>
> it seems to be we can use fls(pf->mask) to replace ctl->soc->mwr_field_width
>

pf->mask cannot be used direct. It needs conversion.For example:
pf->mask: 0x1f -> value_new = (1 << 4) | value;
pf->mask: 0xf -> value_new = (1 << 3) | value;
pf->mask: 0x7 -> value_new = (1 << 2) | value;

The code size of perform conversion is greater than using a direct
mwr_field_width field.


> > + if (value_new == value)
> > + dev_notice(pctl->dev,
> > + "invalid mode 0x%x, use it by ignoring MSBit!\n",
> > + value);
> > + mtk_w32(pctl, pf->index, pf->offset + MWR_OFFSET,
> > + value_new << pf->bitpos);
> > +}
> > +
> > static int mtk_hw_pin_field_lookup(struct mtk_pinctrl *hw,
> > const struct mtk_pin_desc *desc,
> > int field, struct mtk_pin_field *pfd)
> > @@ -197,10 +252,16 @@ int mtk_hw_set_value(struct mtk_pinctrl *hw, const struct mtk_pin_desc *desc,
> > if (value < 0 || value > pf.mask)
> > return -EINVAL;
> >
> > - if (!pf.next)
> > - mtk_rmw(hw, pf.index, pf.offset, pf.mask << pf.bitpos,
> > - (value & pf.mask) << pf.bitpos);
> > - else
> > + if (!pf.next) {
> > + if (hw->soc->race_free_access) {
>
> let's create an extra flags caps under hw->soc and the SoC capability
> check, something like hw->soc->caps & MTK_HW_CAPS_RMW_ATOMIC to easily
> extend various things for future SoC
>
> > + if (field == PINCTRL_PIN_REG_MODE)
> > + mtk_hw_set_mode_race_free(hw, &pf, value);
> > + else
> > + mtk_hw_set_value_race_free(hw, &pf, value);
> > + }
>
> let's create a function holding that specific hardware stuff (at least
> currently it look like), something like
>
> static void mtk_hw_rmw(struct mtk_pinctrl *pctl, struct mtk_pin_field *pf)
> {
> if (pf->field == PINCTRL_PIN_REG_MODE) /* create a member field for pf */
> mtk_hw_mwr(...);
> else
> mtk_hw_w1sc(...);
> }
>

Sine there is no member 'field' in struct mtk_pin_field, pf->field
cannot be used.
Therefore an extra function parameter is required if you want to use a
standalone function mtk_hw_rmw. Like this:

void mtk_hw_rmw(struct mtk_pinctrl *pctl, struct mtk_pin_field *pf,
int field, u32 value)
{
if (field == PINCTRL_PIN_REG_MODE)
mtk_hw_set_mode_race_free(hw, &pf, value);
else
mtk_hw_set_value_race_free(hw, &pf, value);
}

I wonder the necessity/efficiency of such extra intermediate function
with many function parameters.


> > + mtk_rmw(hw, pf.index, pf.offset, pf.mask << pf.bitpos,
> > + (value & pf.mask) << pf.bitpos);
> > + } else
> > mtk_hw_write_cross_field(hw, &pf, value);
> >
> > return 0;
> > diff --git a/drivers/pinctrl/mediatek/pinctrl-mtk-common-v2.h b/drivers/pinctrl/mediatek/pinctrl-mtk-common-v2.h
> > index 27df087..95fb329 100644
> > --- a/drivers/pinctrl/mediatek/pinctrl-mtk-common-v2.h
> > +++ b/drivers/pinctrl/mediatek/pinctrl-mtk-common-v2.h
> > @@ -203,6 +203,8 @@ struct mtk_pin_soc {
> > /* Specific parameters per SoC */
> > u8 gpio_m;
> > bool ies_present;
> > + bool race_free_access;
> > + unsigned int mwr_field_width;
> > const char * const *base_names;
> > unsigned int nbase_names;
> >
> > --
> > 1.8.1.1.dirty