[U-Boot] [PATCH] ARM: socfpga: Put stack at the end of SRAM

[Simon Goldschmidt]

On 14.05.2018 17:51, Stefan Roese wrote:
> On 14.05.2018 11:06, Marek Vasut wrote:
>> On 05/14/2018 11:01 AM, Simon Goldschmidt wrote:
>>>
>>>
>>> On 14.05.2018 10:17, Marek Vasut wrote:
>>>> On 05/14/2018 10:03 AM, Simon Goldschmidt wrote:
>>>>> On 12.05.2018 22:27, Marek Vasut wrote:
>>>>>> The global data are in the .data section, so there's no point in
>>>>>> reserving any space for it above stack. Put stack at the end of
>>>>>> SRAM.
>>>>>>
>>>>>> Signed-off-by: Marek Vasut <marex at denx.de>
>>>>>> Cc: Chin Liang See <chin.liang.see at intel.com>
>>>>>> Cc: Dinh Nguyen <dinguyen at kernel.org>
>>>>>> ---
>>>>>>     include/configs/socfpga_common.h | 4 +---
>>>>>>     1 file changed, 1 insertion(+), 3 deletions(-)
>>>>>>
>>>>>> diff --git a/include/configs/socfpga_common.h
>>>>>> b/include/configs/socfpga_common.h
>>>>>> index 4de2aa7929..cb67d539b1 100644
>>>>>> --- a/include/configs/socfpga_common.h
>>>>>> +++ b/include/configs/socfpga_common.h
>>>>>> @@ -35,10 +35,8 @@
>>>>>>     #define CONFIG_SYS_INIT_RAM_ADDR    0xFFE00000
>>>>>>     #define CONFIG_SYS_INIT_RAM_SIZE    0x40000 /* 256KB */
>>>>>>     #endif
>>>>>> -#define CONFIG_SYS_INIT_SP_OFFSET        \
>>>>>> -    (CONFIG_SYS_INIT_RAM_SIZE - GENERATED_GBL_DATA_SIZE)
>>>>>>     #define CONFIG_SYS_INIT_SP_ADDR            \
>>>>>> -    (CONFIG_SYS_INIT_RAM_ADDR + CONFIG_SYS_INIT_SP_OFFSET)
>>>>>> +    (CONFIG_SYS_INIT_RAM_ADDR + CONFIG_SYS_INIT_RAM_SIZE)
>>>>>
>>>>> I think this interferes with the bootcount being stored at the end of
>>>>> SRAM. See the defconfigs of 'is1' and 'sr1500' boards - or my local
>>>>> defconfig here :-)
>>>>>
>>>>> I'd really like to keep some room here for bootcounter and some 
>>>>> related
>>>>> things I store here.
>>>>
>>>> So if I understand it correctly, before this patch, the boards 
>>>> overwrote
>>>> the global data ?
>>>
>>> Ehrm, reading up on the init code path in the S and C files, it seems
>>> like that, yes. I just used the two boards as an example. Are they
>>> actively maintained? get_maintainer.pl shows the boards are maintained
>>> by denx, so maybe you could clarify this?
>>
>> +CC Stefan.
> 
> Thanks. At the time I added the bootcounter support on this platform,
> it used to work. I might have missed something though. Right now,
> I don't have access to such a platform, so its hard for me to test
> something. If you have some patches to fix any potential overlapping
> memory usages, I'll gladly review them.

I just checked out ae996c80 and from what I can tell, the RAM layout was 
the same at that time: include/generated/generated-asm-offsets.h contains:
- #define GENERATED_GBL_DATA_SIZE 208
- #define GD_SIZE 200

That leaves us with an 8 byte alignment gap at the end of the SRAM and 
that's where the boot counter is located. So the global data is not 
overwritten, but this is more or less by chance, any change to the 
struct definition might change this.

Looking at just another example: the x600 board defines its SRAM_SIZE 8 
bytes smaller than what it really is. That seems like a good idea for 
socfpga, too. I'll see if I can provide a patch.

Slightly off-topic: I have a constellation where the boot-counter is not 
reset although the power has been pulled. This device has big capacitors 
and the SRAM content seems to survive although the processor is reset. 
To me it seems as if looking at the boot reason might be safer than 
checking for a magic word... Any ideas?

Thanks,
Simon

> 
>>> Chances are high that these bytes just fit into the 16-byte alignment
>>> allocated by GENERATED_GBL_DATA_SIZE (and on my board it *is* like 
>>> that).
>>>
>>>>
>>>> Anyway, if you need some extra space, look at
>>>> board_init_f_alloc_reserve() .
>>>>
>>>
>>> OK, but that function does not seem to be overridable? How can I reserve
>>> my own memory here?
>>
>> Add a __weak there I guess ?
>>
>>> And reading this function, I guess your commit message is a bit
>>> misleading. It seems that the purpose of CONFIG_SYS_INIT_SP_OFFSET like
>>> it was was to reserve memory for the initial gd (before relocation),
>>> which is now done in board_init_f_alloc_reserve().
>>
>> The code is confusing as hell, I give you that.
> 
> Thanks,
> Stefan
>