sun9i pll4 upstream kernel code seems wrong

sun9i pll4 upstream kernel code seems wrong

[Hans de Goede]

Hi ChenYu,

Looking at drivers/clk/sunxi/clk-sun9i-core.c:

sun9i_a80_get_pll4_factors(), and comparing it with
the A80 user manual, things seem way off, this seems
more accurate (although also possibly not quite)
for pll1 / pll2 then for pll4, and the comment at
the top does mention PLL1 once.

Note according to the datasheet pll4 should be treated
as an inmutable pll fixed at 960 MHz, so maybe we should
just drop the get_factors function for it ?

Luckily the struct clk_factors_config sun9i_a80_pll4_config
is correct, so as long as we do not try to change the
rate the current upstream code for pll4 does work.

Regards,

Hans

sun9i pll4 upstream kernel code seems wrong

[Chen-Yu Tsai]

Hi,

On Sat, Jan 17, 2015 at 5:34 PM, Hans de Goede <hdegoede@redhat.com> wrote:
> Hi ChenYu,
>
> Looking at drivers/clk/sunxi/clk-sun9i-core.c:
>
> sun9i_a80_get_pll4_factors(), and comparing it with
> the A80 user manual, things seem way off, this seems
> more accurate (although also possibly not quite)
> for pll1 / pll2 then for pll4, and the comment at
> the top does mention PLL1 once.

PLL1 was mentioned as I forgot to change that one in
the comment. I copied the comment section from others.

Other than that, I'm not sure what part is off. The
code was done without the user manual, only with the
SDK bits. I think in the SDK the minimum value of 12
for N factor was not mentioned. But other than that,
it should work fine.

Note the "div" variable uses a base frequency of 6,
not 24. I used a spreadsheet to do some calculations
to see if the code does output the right numbers.
Unfortunately I don't have either the notes or the
spreadsheet I used back then. It would take me a bit
of time to check.

> Note according to the datasheet pll4 should be treated
> as an inmutable pll fixed at 960 MHz, so maybe we should
> just drop the get_factors function for it ?

Is that acceptable? Or is there a readonly flag we should
set for the clock?

Regards,
ChenYu

> Luckily the struct clk_factors_config sun9i_a80_pll4_config
> is correct, so as long as we do not try to change the
> rate the current upstream code for pll4 does work.
>
> Regards,
>
> Hans

sun9i pll4 upstream kernel code seems wrong

[Hans de Goede]

Hi,

On 17-01-15 11:37, Chen-Yu Tsai wrote:
> Hi,
>
> On Sat, Jan 17, 2015 at 5:34 PM, Hans de Goede <hdegoede@redhat.com> wrote:
>> Hi ChenYu,
>>
>> Looking at drivers/clk/sunxi/clk-sun9i-core.c:
>>
>> sun9i_a80_get_pll4_factors(), and comparing it with
>> the A80 user manual, things seem way off, this seems
>> more accurate (although also possibly not quite)
>> for pll1 / pll2 then for pll4, and the comment at
>> the top does mention PLL1 once.
>
> PLL1 was mentioned as I forgot to change that one in
> the comment. I copied the comment section from others.
>
> Other than that, I'm not sure what part is off. The
> code was done without the user manual, only with the
> SDK bits. I think in the SDK the minimum value of 12
> for N factor was not mentioned. But other than that,
> it should work fine.

Ok, so looking at the code again it is not that much
of, just hard to read, and it does contain some errors:

         /* divs above 256 cannot be odd */
         if (div > 256)
                 div = round_up(div, 2);

Should be:

         /* divs above 255 must be a multiple of 2 */
         if (div > 255)
                 div = round_up(div, 2);

Note the 255, and replacing must be odd with
a multiple of 2, as this had nothing to do with odd /
even (more on that later).

Likewise this:

	/* divs above 512 must be a multiple of 4 */
         if (div > 512)
                 div = round_up(div, 4);

Should have s/512/511/ done on it.

And this:

         /* m will be 1 if div is odd */
         if (div & 1)
                 *m = 1;
         else
                 *m = 0;

Is nonsense, the div may have been odd all along,
(never been rounded) and we should still use m = 1, e.g.
to make 78 MHz.

So this should be:

         if (div < 256)
                 *m = 1;
         else
                 *m = 0;

Preferably I would like to see the entire thing rewritten
into something IMHO much more readable / less error prone,
like this:

{
	u8 n;
	u8 m = 1;
	u8 p = 1;

	/* Normalize value to a 6M multiple */
	n = DIV_ROUND_UP(*freq, 6000000);

	if (n > 255) {
		m = 0;
		n = (n + 1) / 2;
	}
		
	if (n > 255) {
		p = 0;
		n = (n + 1) / 2;
	}

	if (n > 255)
		n = 255;
	else if (n < 12)
		n = 12;

	*freq = ((24000000 * n) >> p) / (m + 1);

	if (n_ret == NULL)
		return;

	*n_ret = n;
	*m_ret = m;
	*p_ret = p;
}

With the n / m / p function parameters renamed to foo_ret.

Regards,

Hans


>
> Note the "div" variable uses a base frequency of 6,
> not 24. I used a spreadsheet to do some calculations
> to see if the code does output the right numbers.
> Unfortunately I don't have either the notes or the
> spreadsheet I used back then. It would take me a bit
> of time to check.
>
>> Note according to the datasheet pll4 should be treated
>> as an inmutable pll fixed at 960 MHz, so maybe we should
>> just drop the get_factors function for it ?
>
> Is that acceptable? Or is there a readonly flag we should
> set for the clock?
>
> Regards,
> ChenYu
>
>> Luckily the struct clk_factors_config sun9i_a80_pll4_config
>> is correct, so as long as we do not try to change the
>> rate the current upstream code for pll4 does work.
>>
>> Regards,
>>
>> Hans

sun9i pll4 upstream kernel code seems wrong

[Hans de Goede]

Hi,

On 17-01-15 12:12, Hans de Goede wrote:
> Hi,
>
> On 17-01-15 11:37, Chen-Yu Tsai wrote:
>> Hi,
>>
>> On Sat, Jan 17, 2015 at 5:34 PM, Hans de Goede <hdegoede@redhat.com> wrote:
>>> Hi ChenYu,
>>>
>>> Looking at drivers/clk/sunxi/clk-sun9i-core.c:
>>>
>>> sun9i_a80_get_pll4_factors(), and comparing it with
>>> the A80 user manual, things seem way off, this seems
>>> more accurate (although also possibly not quite)
>>> for pll1 / pll2 then for pll4, and the comment at
>>> the top does mention PLL1 once.
>>
>> PLL1 was mentioned as I forgot to change that one in
>> the comment. I copied the comment section from others.
>>
>> Other than that, I'm not sure what part is off. The
>> code was done without the user manual, only with the
>> SDK bits. I think in the SDK the minimum value of 12
>> for N factor was not mentioned. But other than that,
>> it should work fine.
>
> Ok, so looking at the code again it is not that much
> of, just hard to read, and it does contain some errors:
>
>          /* divs above 256 cannot be odd */
>          if (div > 256)
>                  div = round_up(div, 2);
>
> Should be:
>
>          /* divs above 255 must be a multiple of 2 */
>          if (div > 255)
>                  div = round_up(div, 2);
>
> Note the 255, and replacing must be odd with
> a multiple of 2, as this had nothing to do with odd /
> even (more on that later).
>
> Likewise this:
>
>      /* divs above 512 must be a multiple of 4 */
>          if (div > 512)
>                  div = round_up(div, 4);
>
> Should have s/512/511/ done on it.
>
> And this:
>
>          /* m will be 1 if div is odd */
>          if (div & 1)
>                  *m = 1;
>          else
>                  *m = 0;
>
> Is nonsense, the div may have been odd all along,
> (never been rounded) and we should still use m = 1, e.g.
> to make 78 MHz.

Correction this should read:

Is nonsense, the div may have been *even* all along,
(never been rounded) and we should still use m = 1, e.g.
to make *72* MHz.

Regards,

Hans

>
> So this should be:
>
>          if (div < 256)
>                  *m = 1;
>          else
>                  *m = 0;
>
> Preferably I would like to see the entire thing rewritten
> into something IMHO much more readable / less error prone,
> like this:
>
> {
>      u8 n;
>      u8 m = 1;
>      u8 p = 1;
>
>      /* Normalize value to a 6M multiple */
>      n = DIV_ROUND_UP(*freq, 6000000);
>
>      if (n > 255) {
>          m = 0;
>          n = (n + 1) / 2;
>      }
>
>      if (n > 255) {
>          p = 0;
>          n = (n + 1) / 2;
>      }
>
>      if (n > 255)
>          n = 255;
>      else if (n < 12)
>          n = 12;
>
>      *freq = ((24000000 * n) >> p) / (m + 1);
>
>      if (n_ret == NULL)
>          return;
>
>      *n_ret = n;
>      *m_ret = m;
>      *p_ret = p;
> }
>
> With the n / m / p function parameters renamed to foo_ret.
>
> Regards,
>
> Hans
>
>
>>
>> Note the "div" variable uses a base frequency of 6,
>> not 24. I used a spreadsheet to do some calculations
>> to see if the code does output the right numbers.
>> Unfortunately I don't have either the notes or the
>> spreadsheet I used back then. It would take me a bit
>> of time to check.
>>
>>> Note according to the datasheet pll4 should be treated
>>> as an inmutable pll fixed at 960 MHz, so maybe we should
>>> just drop the get_factors function for it ?
>>
>> Is that acceptable? Or is there a readonly flag we should
>> set for the clock?
>>
>> Regards,
>> ChenYu
>>
>>> Luckily the struct clk_factors_config sun9i_a80_pll4_config
>>> is correct, so as long as we do not try to change the
>>> rate the current upstream code for pll4 does work.
>>>
>>> Regards,
>>>
>>> Hans

sun9i pll4 upstream kernel code seems wrong

[Chen-Yu Tsai]

Hi,

On Sat, Jan 17, 2015 at 7:15 PM, Hans de Goede <hdegoede@redhat.com> wrote:
> Hi,
>
>
> On 17-01-15 12:12, Hans de Goede wrote:
>>
>> Hi,
>>
>> On 17-01-15 11:37, Chen-Yu Tsai wrote:
>>>
>>> Hi,
>>>
>>> On Sat, Jan 17, 2015 at 5:34 PM, Hans de Goede <hdegoede@redhat.com>
>>> wrote:
>>>>
>>>> Hi ChenYu,
>>>>
>>>> Looking at drivers/clk/sunxi/clk-sun9i-core.c:
>>>>
>>>> sun9i_a80_get_pll4_factors(), and comparing it with
>>>> the A80 user manual, things seem way off, this seems
>>>> more accurate (although also possibly not quite)
>>>> for pll1 / pll2 then for pll4, and the comment at
>>>> the top does mention PLL1 once.
>>>
>>>
>>> PLL1 was mentioned as I forgot to change that one in
>>> the comment. I copied the comment section from others.
>>>
>>> Other than that, I'm not sure what part is off. The
>>> code was done without the user manual, only with the
>>> SDK bits. I think in the SDK the minimum value of 12
>>> for N factor was not mentioned. But other than that,
>>> it should work fine.
>>
>>
>> Ok, so looking at the code again it is not that much
>> of, just hard to read, and it does contain some errors:
>>
>>          /* divs above 256 cannot be odd */
>>          if (div > 256)
>>                  div = round_up(div, 2);
>>
>> Should be:
>>
>>          /* divs above 255 must be a multiple of 2 */
>>          if (div > 255)
>>                  div = round_up(div, 2);
>>
>> Note the 255, and replacing must be odd with
>> a multiple of 2, as this had nothing to do with odd /
>> even (more on that later).
>>
>> Likewise this:
>>
>>      /* divs above 512 must be a multiple of 4 */
>>          if (div > 512)
>>                  div = round_up(div, 4);
>>
>> Should have s/512/511/ done on it.
>>
>> And this:
>>
>>          /* m will be 1 if div is odd */
>>          if (div & 1)
>>                  *m = 1;
>>          else
>>                  *m = 0;
>>
>> Is nonsense, the div may have been odd all along,
>> (never been rounded) and we should still use m = 1, e.g.
>> to make 78 MHz.
>
>
> Correction this should read:
>
> Is nonsense, the div may have been *even* all along,
> (never been rounded) and we should still use m = 1, e.g.
> to make *72* MHz.

The m and p constraints were based on comments in the
original SDK code, which weren't very clear as to if
they were hardware constraints or just preferred
behavior in the code path. Looking back, I agree the
work done by me here could have been better.

> Regards,
>
> Hans
>
>
>>
>> So this should be:
>>
>>          if (div < 256)
>>                  *m = 1;
>>          else
>>                  *m = 0;
>>
>> Preferably I would like to see the entire thing rewritten
>> into something IMHO much more readable / less error prone,
>> like this:
>>
>> {
>>      u8 n;
>>      u8 m = 1;
>>      u8 p = 1;
>>
>>      /* Normalize value to a 6M multiple */
>>      n = DIV_ROUND_UP(*freq, 6000000);
>>
>>      if (n > 255) {
>>          m = 0;
>>          n = (n + 1) / 2;
>>      }
>>
>>      if (n > 255) {
>>          p = 0;
>>          n = (n + 1) / 2;
>>      }
>>
>>      if (n > 255)
>>          n = 255;
>>      else if (n < 12)
>>          n = 12;
>>
>>      *freq = ((24000000 * n) >> p) / (m + 1);
>>
>>      if (n_ret == NULL)
>>          return;
>>
>>      *n_ret = n;
>>      *m_ret = m;
>>      *p_ret = p;
>> }
>>
>> With the n / m / p function parameters renamed to foo_ret.

Since you've pretty much wrote the whole function, would
you send a patch out when you get the chance?

Thanks
ChenYu

>> Regards,
>>
>> Hans
>>
>>
>>>
>>> Note the "div" variable uses a base frequency of 6,
>>> not 24. I used a spreadsheet to do some calculations
>>> to see if the code does output the right numbers.
>>> Unfortunately I don't have either the notes or the
>>> spreadsheet I used back then. It would take me a bit
>>> of time to check.
>>>
>>>> Note according to the datasheet pll4 should be treated
>>>> as an inmutable pll fixed at 960 MHz, so maybe we should
>>>> just drop the get_factors function for it ?
>>>
>>>
>>> Is that acceptable? Or is there a readonly flag we should
>>> set for the clock?
>>>
>>> Regards,
>>> ChenYu
>>>
>>>> Luckily the struct clk_factors_config sun9i_a80_pll4_config
>>>> is correct, so as long as we do not try to change the
>>>> rate the current upstream code for pll4 does work.
>>>>
>>>> Regards,
>>>>
>>>> Hans

sun9i pll4 upstream kernel code seems wrong

[Hans de Goede]

Hi,

On 17-01-15 13:09, Chen-Yu Tsai wrote:
> Hi,
>
> On Sat, Jan 17, 2015 at 7:15 PM, Hans de Goede <hdegoede@redhat.com> wrote:
>> Hi,
>>
>>
>> On 17-01-15 12:12, Hans de Goede wrote:
>>>
>>> Hi,
>>>
>>> On 17-01-15 11:37, Chen-Yu Tsai wrote:
>>>>
>>>> Hi,
>>>>
>>>> On Sat, Jan 17, 2015 at 5:34 PM, Hans de Goede <hdegoede@redhat.com>
>>>> wrote:
>>>>>
>>>>> Hi ChenYu,
>>>>>
>>>>> Looking at drivers/clk/sunxi/clk-sun9i-core.c:
>>>>>
>>>>> sun9i_a80_get_pll4_factors(), and comparing it with
>>>>> the A80 user manual, things seem way off, this seems
>>>>> more accurate (although also possibly not quite)
>>>>> for pll1 / pll2 then for pll4, and the comment at
>>>>> the top does mention PLL1 once.
>>>>
>>>>
>>>> PLL1 was mentioned as I forgot to change that one in
>>>> the comment. I copied the comment section from others.
>>>>
>>>> Other than that, I'm not sure what part is off. The
>>>> code was done without the user manual, only with the
>>>> SDK bits. I think in the SDK the minimum value of 12
>>>> for N factor was not mentioned. But other than that,
>>>> it should work fine.
>>>
>>>
>>> Ok, so looking at the code again it is not that much
>>> of, just hard to read, and it does contain some errors:
>>>
>>>           /* divs above 256 cannot be odd */
>>>           if (div > 256)
>>>                   div = round_up(div, 2);
>>>
>>> Should be:
>>>
>>>           /* divs above 255 must be a multiple of 2 */
>>>           if (div > 255)
>>>                   div = round_up(div, 2);
>>>
>>> Note the 255, and replacing must be odd with
>>> a multiple of 2, as this had nothing to do with odd /
>>> even (more on that later).
>>>
>>> Likewise this:
>>>
>>>       /* divs above 512 must be a multiple of 4 */
>>>           if (div > 512)
>>>                   div = round_up(div, 4);
>>>
>>> Should have s/512/511/ done on it.
>>>
>>> And this:
>>>
>>>           /* m will be 1 if div is odd */
>>>           if (div & 1)
>>>                   *m = 1;
>>>           else
>>>                   *m = 0;
>>>
>>> Is nonsense, the div may have been odd all along,
>>> (never been rounded) and we should still use m = 1, e.g.
>>> to make 78 MHz.
>>
>>
>> Correction this should read:
>>
>> Is nonsense, the div may have been *even* all along,
>> (never been rounded) and we should still use m = 1, e.g.
>> to make *72* MHz.
>
> The m and p constraints were based on comments in the
> original SDK code, which weren't very clear as to if
> they were hardware constraints or just preferred
> behavior in the code path. Looking back, I agree the
> work done by me here could have been better.
>
>> Regards,
>>
>> Hans
>>
>>
>>>
>>> So this should be:
>>>
>>>           if (div < 256)
>>>                   *m = 1;
>>>           else
>>>                   *m = 0;
>>>
>>> Preferably I would like to see the entire thing rewritten
>>> into something IMHO much more readable / less error prone,
>>> like this:
>>>
>>> {
>>>       u8 n;
>>>       u8 m = 1;
>>>       u8 p = 1;
>>>
>>>       /* Normalize value to a 6M multiple */
>>>       n = DIV_ROUND_UP(*freq, 6000000);
>>>
>>>       if (n > 255) {
>>>           m = 0;
>>>           n = (n + 1) / 2;
>>>       }
>>>
>>>       if (n > 255) {
>>>           p = 0;
>>>           n = (n + 1) / 2;
>>>       }
>>>
>>>       if (n > 255)
>>>           n = 255;
>>>       else if (n < 12)
>>>           n = 12;
>>>
>>>       *freq = ((24000000 * n) >> p) / (m + 1);
>>>
>>>       if (n_ret == NULL)
>>>           return;
>>>
>>>       *n_ret = n;
>>>       *m_ret = m;
>>>       *p_ret = p;
>>> }
>>>
>>> With the n / m / p function parameters renamed to foo_ret.
>
> Since you've pretty much wrote the whole function, would
> you send a patch out when you get the chance?

Sure, the reason I was proposing this by mail first was to
see if you agree that something like the above will be better,
if you agree I'll turn this into a patch.

Regards,

Hans

sun9i pll4 upstream kernel code seems wrong

[Chen-Yu Tsai]

On Sat, Jan 17, 2015 at 8:30 PM, Hans de Goede <hdegoede@redhat.com> wrote:
> Hi,
>
>
> On 17-01-15 13:09, Chen-Yu Tsai wrote:
>>
>> Hi,
>>
>> On Sat, Jan 17, 2015 at 7:15 PM, Hans de Goede <hdegoede@redhat.com>
>> wrote:
>>>
>>> Hi,
>>>
>>>
>>> On 17-01-15 12:12, Hans de Goede wrote:
>>>>
>>>>
>>>> Hi,
>>>>
>>>> On 17-01-15 11:37, Chen-Yu Tsai wrote:
>>>>>
>>>>>
>>>>> Hi,
>>>>>
>>>>> On Sat, Jan 17, 2015 at 5:34 PM, Hans de Goede <hdegoede@redhat.com>
>>>>> wrote:
>>>>>>
>>>>>>
>>>>>> Hi ChenYu,
>>>>>>
>>>>>> Looking at drivers/clk/sunxi/clk-sun9i-core.c:
>>>>>>
>>>>>> sun9i_a80_get_pll4_factors(), and comparing it with
>>>>>> the A80 user manual, things seem way off, this seems
>>>>>> more accurate (although also possibly not quite)
>>>>>> for pll1 / pll2 then for pll4, and the comment at
>>>>>> the top does mention PLL1 once.
>>>>>
>>>>>
>>>>>
>>>>> PLL1 was mentioned as I forgot to change that one in
>>>>> the comment. I copied the comment section from others.
>>>>>
>>>>> Other than that, I'm not sure what part is off. The
>>>>> code was done without the user manual, only with the
>>>>> SDK bits. I think in the SDK the minimum value of 12
>>>>> for N factor was not mentioned. But other than that,
>>>>> it should work fine.
>>>>
>>>>
>>>>
>>>> Ok, so looking at the code again it is not that much
>>>> of, just hard to read, and it does contain some errors:
>>>>
>>>>           /* divs above 256 cannot be odd */
>>>>           if (div > 256)
>>>>                   div = round_up(div, 2);
>>>>
>>>> Should be:
>>>>
>>>>           /* divs above 255 must be a multiple of 2 */
>>>>           if (div > 255)
>>>>                   div = round_up(div, 2);
>>>>
>>>> Note the 255, and replacing must be odd with
>>>> a multiple of 2, as this had nothing to do with odd /
>>>> even (more on that later).
>>>>
>>>> Likewise this:
>>>>
>>>>       /* divs above 512 must be a multiple of 4 */
>>>>           if (div > 512)
>>>>                   div = round_up(div, 4);
>>>>
>>>> Should have s/512/511/ done on it.
>>>>
>>>> And this:
>>>>
>>>>           /* m will be 1 if div is odd */
>>>>           if (div & 1)
>>>>                   *m = 1;
>>>>           else
>>>>                   *m = 0;
>>>>
>>>> Is nonsense, the div may have been odd all along,
>>>> (never been rounded) and we should still use m = 1, e.g.
>>>> to make 78 MHz.
>>>
>>>
>>>
>>> Correction this should read:
>>>
>>> Is nonsense, the div may have been *even* all along,
>>> (never been rounded) and we should still use m = 1, e.g.
>>> to make *72* MHz.
>>
>>
>> The m and p constraints were based on comments in the
>> original SDK code, which weren't very clear as to if
>> they were hardware constraints or just preferred
>> behavior in the code path. Looking back, I agree the
>> work done by me here could have been better.
>>
>>> Regards,
>>>
>>> Hans
>>>
>>>
>>>>
>>>> So this should be:
>>>>
>>>>           if (div < 256)
>>>>                   *m = 1;
>>>>           else
>>>>                   *m = 0;
>>>>
>>>> Preferably I would like to see the entire thing rewritten
>>>> into something IMHO much more readable / less error prone,
>>>> like this:
>>>>
>>>> {
>>>>       u8 n;
>>>>       u8 m = 1;
>>>>       u8 p = 1;
>>>>
>>>>       /* Normalize value to a 6M multiple */
>>>>       n = DIV_ROUND_UP(*freq, 6000000);
>>>>
>>>>       if (n > 255) {
>>>>           m = 0;
>>>>           n = (n + 1) / 2;
>>>>       }
>>>>
>>>>       if (n > 255) {
>>>>           p = 0;
>>>>           n = (n + 1) / 2;
>>>>       }
>>>>
>>>>       if (n > 255)
>>>>           n = 255;
>>>>       else if (n < 12)
>>>>           n = 12;
>>>>
>>>>       *freq = ((24000000 * n) >> p) / (m + 1);
>>>>
>>>>       if (n_ret == NULL)
>>>>           return;
>>>>
>>>>       *n_ret = n;
>>>>       *m_ret = m;
>>>>       *p_ret = p;
>>>> }
>>>>
>>>> With the n / m / p function parameters renamed to foo_ret.
>>
>>
>> Since you've pretty much wrote the whole function, would
>> you send a patch out when you get the chance?
>
>
> Sure, the reason I was proposing this by mail first was to
> see if you agree that something like the above will be better,
> if you agree I'll turn this into a patch.

I say go for it. I'm a bit busy with travel preparations and
other stuff this week, so I haven't tested the sample above.


ChenYu