Re: [PATCH v2 2/2] i2c: mediatek: Add i2c ac-timing adjust support

[Geert Uytterhoeven]

Hi Qii,

On Wed, May 20, 2020 at 10:44 AM Qii Wang <qii.wang@xxxxxxxxxxxx> wrote:
> On Tue, 2020-05-19 at 09:14 +0200, Geert Uytterhoeven wrote:
> > On Tue, May 19, 2020 at 4:59 AM Qii Wang <qii.wang@xxxxxxxxxxxx> wrote:
> > > On Mon, 2020-05-18 at 17:44 +0200, Geert Uytterhoeven wrote:
> > > > On Thu, May 14, 2020 at 3:13 PM Qii Wang <qii.wang@xxxxxxxxxxxx> wrote:
> > > > > This patch adds a algorithm to calculate some ac-timing parameters
> > > > > which can fully meet I2C Spec.
> > > > >
> > > > > Signed-off-by: Qii Wang <qii.wang@xxxxxxxxxxxx>
> > > > > ---
> > > > >  drivers/i2c/busses/i2c-mt65xx.c | 328 +++++++++++++++++++++++++++++++++-------
> > > > >  1 file changed, 277 insertions(+), 51 deletions(-)
> > > > >
> > > > > diff --git a/drivers/i2c/busses/i2c-mt65xx.c b/drivers/i2c/busses/i2c-mt65xx.c
> > > > > index 0ca6c38a..7020618 100644
> > > > > --- a/drivers/i2c/busses/i2c-mt65xx.c
> > > > > +++ b/drivers/i2c/busses/i2c-mt65xx.c
> > > >
> > > > > +/*
> > > > > + * Check and Calculate i2c ac-timing
> > > > > + *
> > > > > + * Hardware design:
> > > > > + * sample_ns = (1000000000 * (sample_cnt + 1)) / clk_src
> > > > > + * xxx_cnt_div =  spec->min_xxx_ns / sample_ns
> > > > > + *
> > > > > + * Sample_ns is rounded down for xxx_cnt_div would be greater
> > > > > + * than the smallest spec.
> > > > > + * The sda_timing is chosen as the middle value between
> > > > > + * the largest and smallest.
> > > > > + */
> > > > > +static int mtk_i2c_check_ac_timing(struct mtk_i2c *i2c,
> > > > > +                                  unsigned int clk_src,
> > > > > +                                  unsigned int check_speed,
> > > > > +                                  unsigned int step_cnt,
> > > > > +                                  unsigned int sample_cnt)
> > > > > +{
> > > > > +       const struct i2c_spec_values *spec;
> > > > > +       unsigned int su_sta_cnt, low_cnt, high_cnt, max_step_cnt;
> > > > > +       unsigned int sda_max, sda_min, clk_ns, max_sta_cnt = 0x3f;
> > > > > +       long long sample_ns = (1000000000 * (sample_cnt + 1)) / clk_src;
> > > >
> > > > So sample_ns is a 64-bit value. Is that really needed?
> > > >
> > >
> > > (1000000000 * (sample_cnt + 1)) / clk_src value is a 32-bit, (1000000000
> > > * (sample_cnt + 1)) will over 32-bit if sample_cnt is 7.
> >
> > The intermediate value will indeed not fit in 32-bit.
> > But that doesn't mean the end result won't fit in 32-bit.
> > As you divide spec->min_low_ns and spec->min_su_dat_ns (which I assume
> > are small numbers) by sample_ns below, sample_ns cannot be very large,
> > or the quotient will be zero anyway.
> > So just doing the multiplication in 64-bit, followed by a 64-by-32
> > division is probably fine:
> >
> >     unsigned int sample_ns = div_u64(1000000000ULL * (sample_cnt + 1), clk_src);
> >
> > You may want to take precautions for the case where the passed value of
> > clk_src is a small number (can that happen?).
> >
> > BTW, clk_get_rate() returns "unsigned long", while mtk_i2c_set_speed()
> > takes an "unsigned int" parent_clk, which may cause future issues.
> > You may want to change that to "unsigned long", along the whole
> > propagation path, and use div64_ul() instead of div_u64() above.
> >
>
> The return type of div_u64 is u64(unsigned long long), there is a
> compulsory type conversion operator. Do you think it is needed?

The result of a 64-by-32 bit division may indeed not fit in 32-bit, so that's
why it returns u64.
If you know the quotient will always fit, it's fine.

> BTW, we just need to change the type of sample_ns to unsigned int, no
> matter which method is used, what is your opinion?

Indeed.

BTW, I just realize

    long long sample_ns = (1000000000 * (sample_cnt + 1)) / clk_src;

wasn't doing what you wanted anyway, as 1000000000 is (implicit) int,
and sample_cnt is unsigned int, so the multiplication was done in 32-bit,
possible leading to a truncation.  Hence that division was done in 32-bit, too,
that's why I didn't notice a call to __udivdi3() in the assembler output here.

So you have to force the multiplication to be done in 64-bit, e.g.
by changing the constant to 1000000000ULL, and use div_u64() for
the division.

>
> > > I think 1000000000 and clk_src is too big, maybe I can reduce then with
> > > be divided all by 1000.
> > > example:
> > >
> > > unsigned int sample_ns;
> > > unsigned int clk_src_khz = clk_src / 1000;
> >
> > That may cause too much loss of precision.
> >
>
> clk_src is more than MHz and less than GHZ for MTK i2c controller, so it
> wouldn't cause too much loss of precision.

OK, so that would work, too.

> > >
> > > if(clk_src_khz)
> > >         sample_ns = (1000000 * (sample_cnt + 1)) / clk_src_khz;
> > > else
> > >         return -EINVAL;
> > >
> > > > > +       if (!i2c->dev_comp->timing_adjust)
> > > > > +               return 0;
> > > > > +
> > > > > +       if (i2c->dev_comp->ltiming_adjust)
> > > > > +               max_sta_cnt = 0x100;
> > > > > +
> > > > > +       spec = mtk_i2c_get_spec(check_speed);
> > > > > +
> > > > > +       if (i2c->dev_comp->ltiming_adjust)
> > > > > +               clk_ns = 1000000000 / clk_src;
> > > > > +       else
> > > > > +               clk_ns = sample_ns / 2;
> > > > > +
> > > > > +       su_sta_cnt = DIV_ROUND_UP(spec->min_su_sta_ns, clk_ns);
> > > > > +       if (su_sta_cnt > max_sta_cnt)
> > > > > +               return -1;
> > > > > +
> > > > > +       low_cnt = DIV_ROUND_UP(spec->min_low_ns, sample_ns);
> > > >
> > > > So this is a 32-bit by 64-bit division (indeed, not 64-by-32!)

Gr{oetje,eeting}s,

                        Geert

-- 
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@xxxxxxxxxxxxxx

In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
                                -- Linus Torvalds