Re: [PATCH 2/2] clk: Add fractional scale clock support

[Hoan Tran]

Hi Geert,

On Thu, Jun 16, 2016 at 11:43 PM, Geert Uytterhoeven
<geert@xxxxxxxxxxxxxx> wrote:
> On Fri, Jun 17, 2016 at 1:40 AM, Hoan Tran <hotran@xxxxxxx> wrote:
>> This patch adds fractional scale clock support.
>> Fractional scale clock is implemented for a single register field.
>>   Output rate = parent_rate * scale / denominator
>> For example, for 1 / 8 fractional scale, denominator will be 8 and scale
>> will be computed and programmed accordingly.
>>
>> Signed-off-by: Hoan Tran <hotran@xxxxxxx>
>> Signed-off-by: Loc Ho <lho@xxxxxxx>
>
>> --- /dev/null
>> +++ b/drivers/clk/clk-fractional-scale.c
>
>> +static long clk_fs_round_rate(struct clk_hw *hw, unsigned long rate,
>> +                             unsigned long *parent_rate)
>> +{
>> +       struct clk_fractional_scale *fd = to_clk_sf(hw);
>> +       u64 ret, scale;
>> +
>> +       if (!rate || rate >= *parent_rate)
>> +               return *parent_rate;
>> +
>> +       /* freq = parent_rate * scaler / denom */
>> +       ret = rate * fd->denom;
>
> Can this overflow? No, because fd->denom is u64.
> But if fd->denom is changed to u32, it needs a cast to u64 here.
>
>> +       scale = ret / *parent_rate;
>
> As detected by the kbuild test robot, this should use do_div().

Yes, my mistake. I'll fix it.

>
>> +
>> +       ret = (u64) *parent_rate * scale;
>> +       do_div(ret, fd->denom);
>> +
>> +       return ret;
>> +}
>> +
>> +static int clk_fs_set_rate(struct clk_hw *hw, unsigned long rate,
>> +                          unsigned long parent_rate)
>> +{
>> +       struct clk_fractional_scale *fd = to_clk_sf(hw);
>> +       unsigned long flags = 0;
>> +       u64 scale, ret;
>> +       u32 val;
>> +
>> +       /*
>> +        * Compute the scaler:
>> +        *
>> +        * freq = parent_rate * scaler / denom, or
>> +        * scaler = freq * denom / parent_rate
>> +        */
>> +       ret = rate * fd->denom;
>
> Can this overflow? No, because fd->denom is u64.
> But if fd->denom is changed to u32, it needs a cast to u64 here.
>
>> +       scale = ret / (u64)parent_rate;
>
> Don't cast the divider to u64, as this will turn a 64-by-32 division into a
> 64-by-64 division.
> As detected by the kbuild test robot, this should use do_div().

Will fix it.

Thanks
Hoan

>
>> --- a/include/linux/clk-provider.h
>> +++ b/include/linux/clk-provider.h
>
>> +struct clk_fractional_scale {
>> +       struct clk_hw   hw;
>> +       void __iomem    *reg;
>> +       u8              shift;
>> +       u32             mask;
>> +       u64             denom;
>
> u64 sounds overkill to me, but it looks like that was done to avoid overflow in
> the multiplications?
>
>> +       u32             flags;
>> +       spinlock_t      *lock;
>> +};
>
> 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