[U-Boot] SPL broken on i.mx31 platforms

[Benoît Thébaudeau]

)Dear Helmut Raiger,

On Wed, Jul 2, 2014 at 9:04 AM, Helmut Raiger <helmut.raiger at hale.at> wrote:
>>      the commit 41623c91 breaks the SPL on i.mx31 platforms.

Here, you are talking about mx31pdk, right?

>> The original startup code (start.S) was position independent to
>> allow relocation in board_init_f. This is necessary as the internal
>> RAM used by the IPL to load the first 2kB from NAND is also
>> used by the NAND controller to buffer pages.
>> As far as the issue goes: where and how exactly is the code not
>> position independent now?
>
> The first statement is now:
>
>     ldr pc, _reset
> _reset: .word reset
>
> reset is an address generated by the linker (here 0x87dc0060).
> Originally this was:
>
>     b reset
>
> which is a relative branch (here 0x60 bytes forward).
> I have to add, that changing this back does not fix the issue, something
> else is in disorder.
> I'm still working on it.

Do you mean reverting this commit, or only these two lines? In the
former case, where does bisect lead to, and in the latter case, have
you checked the assembler code in the linked SPL for this branch?

>>>> Does changing the startup code back to PIC generate any
>>>> drawbacks on other ARM platforms?
>>
>> How would this changing back to PIC be implemented? Not by reverting
>> the patch, I hope.
>
> No, it will only be a minor change, I think, but I thought there might
> have been an additional intention behind the change to position
> dependent code. One could link the first part to 0xB8000000
> (the original position of the SPL when loaded by the IPL) and
> the part after the relocation to CONFIG_SPL_TEXT_BASE.

Actually, the ROM bootloader first copies the first NAND page to
0xB8000000. Then, the SPL placed here but linked at
CONFIG_SPL_TEXT_BASE copies itself to CONFIG_SPL_TEXT_BASE in order to
free the NFC buffer so that it can be used by the SPL. There is no
relocation going on at this stage, but only a copy, and the SPL code
size is limited to 2 kiB. Then, the SPL does its NAND load job towards
CONFIG_SYS_TEXT_BASE and starts executing the non-SPL binary, which
then relocates itself according to the heap size, etc.

Hence, as you said the code launching board_init_f() must be either
PIC or linked at 0xB8000000, and the rest of the SPL code must be
either PIC or linked at CONFIG_SPL_TEXT_BASE. So the fact that this
commit breaks the "PICness" of the initial branch is indeed an issue.

As a test, can you try to replace the 2 _reset lines in vectors.S with
"b _reset" after "_start:", and a "_reset:" label at the end of the
file? This assumes that start.S starts with the "reset:" label, and
that the start.o .text sections are placed right after the .vectors
section in the SPL linker script, both being true for mx31pdk.

Best regards,
Benoît