Re: [PATCH v1] mmc: mediatek: Fixed bug where clock frequency could be set wrong

[Daniel Kurtz]

On Wed, Mar 1, 2017 at 5:56 PM, Yong Mao <yong.mao@xxxxxxxxxxxx> wrote:
> On Tue, 2017-02-28 at 14:56 +0800, Daniel Kurtz wrote:
>> On Fri, Feb 24, 2017 at 5:38 PM, Yong Mao <yong.mao@xxxxxxxxxxxx> wrote:
>> > From:   Yong Mao <yong.mao@xxxxxxxxxxxx>
>> > To:     Daniel Kurtz <djkurtz@xxxxxxxxxxxx>
>> > Subject:        Re: [PATCH v1] mmc: mediatek: Fixed bug where clock frequency
>> > could be set wrong
>> > Date:   Fri, 24 Feb 2017 17:33:37 +0800
>> >
>> >
>> > On Fri, 2017-02-24 at 17:52 +0900, Daniel Kurtz wrote:
>> >> On Fri, Feb 24, 2017 at 5:22 PM, Yong Mao <yong.mao@xxxxxxxxxxxx>
>> > wrote:
>> >> >
>> >> > From: yong mao <yong.mao@xxxxxxxxxxxx>
>> >> >
>> >> > This patch can fix two issues:
>> >> >
>> >> > Issue 1:
>> >> > The maximum value of clock divider is 0xff.
>> >> > Because the type of div is u32, div may be larger than max_div.
>> >> > In this case, we should use max_div to set the clock frequency.
>> >> >
>> >> > Issue 2:
>> >> > In previous code, we can not set the correct clock frequency when
>> >> > div equals 0xff.
>> >> >
>> >> > Signed-off-by: Yong Mao <yong.mao@xxxxxxxxxxxx>
>> >> > Signed-off-by: Chaotian Jing <chaotian.jing@xxxxxxxxxxxx>
>> >> > ---
>> >> >  drivers/mmc/host/mtk-sd.c |   13 ++++++++++++-
>> >> >  1 file changed, 12 insertions(+), 1 deletion(-)
>> >> >
>> >> > diff --git a/drivers/mmc/host/mtk-sd.c b/drivers/mmc/host/mtk-sd.c
>> >> > index 07f3236..3174445 100644
>> >> > --- a/drivers/mmc/host/mtk-sd.c
>> >> > +++ b/drivers/mmc/host/mtk-sd.c
>> >> > @@ -540,6 +540,7 @@ static void msdc_set_mclk(struct msdc_host
>> > *host, unsigned char timing, u32 hz)
>> >> >         u32 mode;
>> >> >         u32 flags;
>> >> >         u32 div;
>> >> > +       u32 max_div;
>> >>
>> >> There's really no need for this variable.  Just use 0xff below.
>> > For all of our IC, max_div is not a constant.
>> > We will upstream another patch which max_div will get the different
>> > value depending on the IC.
>> > Therefore, we keep the max_div as a variable here.
>>
>> Please add the variable in the patch that uses it as a variable.
>>
>> >
>> >>
>> >> >         u32 sclk;
>> >> >
>> >> >         if (!hz) {
>> >> > @@ -590,8 +591,18 @@ static void msdc_set_mclk(struct msdc_host
>> > *host, unsigned char timing, u32 hz)
>> >> >                         sclk = (host->src_clk_freq >> 2) / div;
>> >> >                 }
>> >> >         }
>> >> > +
>> >> > +       /**
>> >> > +        * The maximum value of div is 0xff.
>> >> > +        * Check if the div is larger than max_div.
>> >> > +        */
>> >> > +       max_div = 0xff;
>> >> > +       if (div > max_div) {
>> >> > +               div = max_div;
>> >> > +               sclk = (host->src_clk_freq >> 2) / div;
>> >> > +       }
>> >> >         sdr_set_field(host->base + MSDC_CFG, MSDC_CFG_CKMOD |
>> > MSDC_CFG_CKDIV,
>> >> > -                       (mode << 8) | (div % 0xff));
>> >> > +                     (mode << 8) | div);
>> >>
>> >> Hmm, I don't know much about this sub-system, but should we even be
>> >> allowing requests to set a frequency that we can't actually achieve
>> >> with the divider?
>> >>
>> >
>> > No. We can not get a frequency that we can't actually achieve with the
>> > divider. This patch is to solve this kind of issue.
>>
>> Sorry, I am trying to understand why we need this patch.
>>
>> AFAICT, it looks like sometimes msdc_set_mclk() is being called with
>> hz that cannot be generated by your hardware.  In particular,
>> sometimes the original code computes "div > 255".
>> To work around this problem, this patch just caps the divider value to
>> 255, which is the maximum divider provided by the hardware.  However,
>> presumably this means that in this case we won't actually be
>> generating the requested hz value.
>>
>> So, can you please explain in what exact scenario this patch is
>> required, and justify why it is ok to generate a clock other than the
>> requested in this case?
>>
>> -Dan
>>
>
> This issue is hidden deeply. It is a boundary related issue.
> Let me take the real value to explain how this issue occurs.
>
> 1. mmc->f_min = host->src_clk_freq / (4 * 255);
>    mmc->f_min = 400000000 / (4 * 255) = 392156;
> 2. mmc core tries to initialize emmc by using 400000hz.
>    If the first try failed, it will retry by using max(300000hz,
> mmc->f_min)= 392156hz
> 3. msdc_set_mclk will be invoked by mmc core to set the clock.
>    and then following code will be executed.
> div = (host->src_clk_freq + ((hz << 2) - 1)) / (hz << 2);
> (div = (400000000 + (392156 << 2) - 1)) / (392156 << 2) = 256)
>
> Why do we use (host->src_clk_freq + ((hz << 2) - 1))?
> ==> Because if we can not get a proper clock frequency, we should set a
> clock frequency which is less than proper clock frequency.
>
> 4. In this IC, it only use 8 bits to indicate the value of clock
> divider. Therefore, 256 is overflow, it is equal 0 here.
> And then we will get a wrong 100000000Hz clock frequency.
>
> Can you understand how this issue occurs now?

Thanks for the excellent explanation.  Things are a lot clearer now.

I think this can be solved even easier by computing f_min the same way:

  mmc->f_min = DIV_ROUND_UP(400000000, 4 * 255);  /* = 392157 */


>
>> >
>> >
>> >> >         sdr_set_bits(host->base + MSDC_CFG, MSDC_CFG_CKPDN);
>> >> >         while (!(readl(host->base + MSDC_CFG) & MSDC_CFG_CKSTB))
>> >> >                 cpu_relax();
>> >> > --
>> >> > 1.7.9.5
>> >> >
>> >> >
>> >> > _______________________________________________
>> >> > Linux-mediatek mailing list
>> >> > Linux-mediatek@xxxxxxxxxxxxxxxxxxx
>> >> > http://lists.infradead.org/mailman/listinfo/linux-mediatek
>> >
>> >
>> >
>
>