[U-Boot] [PATCH v3] regmap: add support for address cell 2

[Simon Glass]

Hi Kever,

On 7 May 2017 at 19:45, Kever Yang <kever.yang at rock-chips.com> wrote:
> Hi Heiko,
>
>
> Thanks for your comments.
>
>
> On 05/05/2017 09:10 PM, Heiko Stuebner wrote:
>>
>> Hi Kever,
>>
>> Am Freitag, 5. Mai 2017, 10:39:35 CEST schrieb Kever Yang:
>>>
>>> ARM64 is using 64bit address which address cell is 2 instead of 1,
>>> update to support it when of-platdata enabled.
>>>
>>> Signed-off-by: Kever Yang <kever.yang at rock-chips.com>
>>
>> This helps make OF_PLATDATA work on my firefly-rk3399 so yay :-),
>> but I don't think it's that easy to solve, see below:
>>
>>> ---
>>>
>>> Changes in v3:
>>> - move of_plat_get_number() into lib/of_plat.c
>>>
>>> Changes in v2:
>>> - rename the fdtdec_get_number() to of_plat_get_number()
>>>
>>>   drivers/core/regmap.c |  9 +++++++++
>>>   include/of_plat.h     | 22 ++++++++++++++++++++++
>>>   lib/Makefile          |  3 +++
>>>   lib/of_plat.c         | 17 +++++++++++++++++
>>>   4 files changed, 51 insertions(+)
>>>   create mode 100644 include/of_plat.h
>>>   create mode 100644 lib/of_plat.c
>>>
>>> diff --git a/drivers/core/regmap.c b/drivers/core/regmap.c
>>> index 3bec3df..c03279e 100644
>>> --- a/drivers/core/regmap.c
>>> +++ b/drivers/core/regmap.c
>>> @@ -12,6 +12,7 @@
>>>   #include <malloc.h>
>>>   #include <mapmem.h>
>>>   #include <regmap.h>
>>> +#include <of_plat.h>
>>>     #include <asm/io.h>
>>>   @@ -49,11 +50,19 @@ int regmap_init_mem_platdata(struct udevice *dev,
>>> u32 *reg, int count,
>>>         if (!map)
>>>                 return -ENOMEM;
>>>   +#ifdef CONFIG_PHYS_64BIT
>>> +       map->base = of_plat_get_number(reg, 2);
>>> +       for (range = map->range; count > 0; reg += 4, range++, count--) {
>>> +               range->start = of_plat_get_number(reg, 2);
>>> +               range->size = of_plat_get_number(reg + 2, 2);
>>> +       }
>>
>> I may just be missing something, but how can you be sure that the
>> cell-size
>> is always 2?
>>
>> For example, there were discussions about 64bit platforms not really
>> needing to add all the 0x0 elements, when the whole io-registers are well
>> below the 4GB mark and for example at least one sunxi also uses this, see
>> for example:
>> allwinner/sun50i-a64.dtsi, altera/socfpga_stratix10.dtsi,
>> broadcom/bcm283x.dtsi and problably more using #address-cells = <1>,
>> #size-cells = <1> for their memory mapped io.
>>
>> And from what I've seen dtoc simply converts the reg property and just
>> ignores #address-cells and #size-cells (or I'm overlooking something).
>>
>> Possible solutions that come to mind would be make dtoc also convert
>> #address-cells and #size-cells, making regmap and everybody check it
>> or alternatively make dtoc convert regs to cell-size 2 in all cases when
>> CONFIG_PHYS_64BIT is set.
>
>
> @Simon, could you take a look about this issue, I really not good at dtoc
> and libfdt, I think maybe we can re-use fdtdec_get_number() if dtoc do not
> do the fdt32_to_cpu()  int dtoc?

I think the best solution is the second one above. However dtoc is a
bit of a pain to change at the moment. I've sent a series to clean it
up.

I'll hopefully take a look at the above in the next few weeks.

Regards,
Simon