Hi Geert,
On Tue, 2020-05-19 at 09:14 +0200, Geert Uytterhoeven wrote:
> Hi Qii,
>
> 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?
BTW, we just need to change the type of sample_ns to unsigned int, no
matter which method is used, what is your opinion?
> > 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.
> >
> > 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
>
