Re: [PATCH v15 06/16] of/fdt: add helper functions for handling properties

[Frank Rowand]

On 10/09/18 18:14, AKASHI, Takahiro wrote:
> Frank, Rob,
> 
> On Tue, Oct 09, 2018 at 10:47:15AM -0700, Frank Rowand wrote:
>> On 10/08/18 17:37, AKASHI, Takahiro wrote:
>>> On Fri, Oct 05, 2018 at 08:23:57AM -0500, Rob Herring wrote:
>>>> On Thu, Oct 4, 2018 at 10:07 PM AKASHI, Takahiro
>>>> <takahiro.akashi@xxxxxxxxxx> wrote:
>>>>>
>>>>> Rob,
>>>>>
>>>>> # I haven't replied to this comment yet.
>>>>>
>>>>> On Fri, Sep 28, 2018 at 08:44:42AM -0500, Rob Herring wrote:
>>>>>> +David Gibson
>>>>>>
>>>>>> On Fri, Sep 28, 2018 at 1:48 AM AKASHI Takahiro
>>>>>> <takahiro.akashi@xxxxxxxxxx> wrote:
>>>>>>>
>>>>>>> These functions will be used later to handle kexec-specific properties
>>>>>>> in arm64's kexec_file implementation.
>>>>>>>
>>>>>>> Signed-off-by: AKASHI Takahiro <takahiro.akashi@xxxxxxxxxx>
>>>>>>> Cc: Rob Herring <robh+dt@xxxxxxxxxx>
>>>>>>> Cc: Frank Rowand <frowand.list@xxxxxxxxx>
>>>>>>> Cc: devicetree@xxxxxxxxxxxxxxx
>>>>>>> ---
>>>>>>>  drivers/of/fdt.c       | 56 ++++++++++++++++++++++++++++++++++++++++++
>>>>>>>  include/linux/of_fdt.h |  4 +++
>>>>>>>  2 files changed, 60 insertions(+)
>>>>>>>
>>>>>>> diff --git a/drivers/of/fdt.c b/drivers/of/fdt.c
>>>>>>> index 800ad252cf9c..c65c31562ccb 100644
>>>>>>> --- a/drivers/of/fdt.c
>>>>>>> +++ b/drivers/of/fdt.c
>>>>>>> @@ -25,6 +25,7 @@
>>>>>>>  #include <linux/debugfs.h>
>>>>>>>  #include <linux/serial_core.h>
>>>>>>>  #include <linux/sysfs.h>
>>>>>>> +#include <linux/types.h>
>>>>>>>
>>>>>>>  #include <asm/setup.h>  /* for COMMAND_LINE_SIZE */
>>>>>>>  #include <asm/page.h>
>>>>>>> @@ -1323,3 +1324,58 @@ late_initcall(of_fdt_raw_init);
>>>>>>>  #endif
>>>>>>>
>>>>>>>  #endif /* CONFIG_OF_EARLY_FLATTREE */
>>>>>>> +
>>>>>>> +#define FDT_ALIGN(x, a)        (((x) + (a) - 1) & ~((a) - 1))
>>>>>>> +#define FDT_TAGALIGN(x)        (FDT_ALIGN((x), FDT_TAGSIZE))
>>>>>>> +
>>>>>>> +int fdt_prop_len(const char *prop_name, int len)
>>>>>>> +{
>>>>>>> +       return (strlen(prop_name) + 1) +
>>>>>>> +               sizeof(struct fdt_property) +
>>>>>>> +               FDT_TAGALIGN(len);
>>>>>>
>>>>>> Looks like you are using this to calculate how much space you need to
>>>>>> allocate in addition to the current DTB for a couple of new or
>>>>>> replaced properties. I'm not sure that this calculation is completely
>>>>>> accurate. And it is strange there doesn't seem to be any libfdt
>>>>>> function for this already. It would be simpler to just add some fixed
>>>>>> additional amount.
>>>>>>
>>>>>> Maybe David G has comments on this?
>>>>>>
>>>>>>> +}
>>>>>>> +
>>>>>>
>>>>>> The rest of this should go in drivers/of/fdt_address.c. Ultimately, it
>>>>>> should go into libfdt, but I'm fine with having it in the kernel for
>>>>>> now.
>>>>>
>>>>> I'd like to have this function in the kernel for now.
>>>>>
>>>>>>> +static void fill_property(void *buf, u64 val64, int cells)
>>>>>>> +{
>>>>>>> +       __be32 val32;
>>>>>>> +
>>>>>>> +       while (cells) {
>>>>>>> +               val32 = cpu_to_fdt32((val64 >> (32 * (--cells))) & U32_MAX);
>>>>>>> +               memcpy(buf, &val32, sizeof(val32));
>>>>>>> +               buf += sizeof(val32);
>>>>>>
>>>>>> This is kind of hard to read. I would copy u-boot's fdt_pack_reg function.
>>>>>
>>>>> Are you sure?
>>>>> I originally implemented this function in a similar way that fdt_pack_reg()
>>>>> was, but, you suggested, in your past comment[1], that we'd be better to
>>>>> have of_read_number()-like implementation.
>>>>>
>>>>> [1] http://lists.infradead.org/pipermail/linux-arm-kernel/2018-May/579118.html
>>>>
>>>> Yeah, you're right. Plus, I'm not sure the u-boot one would work for
>>>> unaligned accesses with armv5 and earlier h/w.
>>>>
>>>> My only comment then is I think you can drop the U32_MAX masking.
>>>
>>> Okay, then I will leave this function, yet renaming it to
>>> cpu64_to_fdt_cells() after Frank's comment.
>>
>> I have second guessed myself and do not like the name I suggested
>> because what the function really does is either cpu32 to be32 or
>> cpu64 to be64.
> 
> Okay.
> 
>> I agree with Rob that readability is important here.  Instead of
>> having a fill_property() function, how about having inline code,
>> something like (untested even for thinkos):
>>
>> 	prop = buf;
>>
>> 	if (addr_cells == 1) {
>> 		*(__be32 *)prop = cpu32_to_be32(addr);
>> 		prop += 4;
>> 	} else {
>> 		*(__be64 *)prop = cpu64_to_be64(addr);
>> 		prop += 8;
>> 	}
>>
>> 	if (size_cells == 1)
>> 		*(__be32 *)prop = cpu32_to_be32(size);
>> 	else
>> 		*(__be64 *)prop = cpu64_to_be64(size);
>>
>> You might want to also give Rob a chance to bike shed on this
>> suggestion.
> 
> Basically, I don't care either way, but what Rob suggested
> is that some architecture(s) might not handle correctly
> unaligned memory access here.

Good point about unaligned memory access.  That rules out my
suggestion above.  But if using the fill_property() implementation,
either come up with a name that better describes what the function
does or add a function header comment that gives the reader a
clue as to what the function accomplishes.

-Frank


> I just want to stay tuned with Rob.
> 
> Thanks,
> -Takahiro Akashi
> 
>> -Frank
>>
>>>
>>> Thanks,
>>> -Takahiro Akashi
>>>>
>>>> Rob
>>>
>>
>