[U-Boot] [PATCH 1/1] spl: u-boot-spl-nodtb.bin: remove .got and .got.plt sections

[Bin Meng]

Hi Heinrich,

On Tue, Sep 12, 2017 at 8:50 AM, Heinrich Schuchardt <xypron.glpk at gmx.de> wrote:
> On 09/12/2017 01:19 AM, Dr. Philipp Tomsich wrote:
>>> On 12 Sep 2017, at 00:40, Heinrich Schuchardt <xypron.glpk at gmx.de> wrote:
>>>
>>> For qemu-x86_64_defconfig on Debian Stretch
>>> ld 2.28 creates a .got and a .got.plt section in
>>> u-boot-spl.
>>>
>>> If we do not ignore these sections objcopy generates an
>>> u-boot-spl-nodtb.bin file of 4 GiB which leads to
>>> subsequent failure of the build process.
>>>
>>> According to
>>> https://sourceware.org/bugzilla/show_bug.cgi?id=22120#c2
>>> objcopy can not generate files with holes.
>>>
>>> With the patch we can run qemu with the generated u-boot.rom.
>>>
>>> Signed-off-by: Heinrich Schuchardt <xypron.glpk at gmx.de>
>>> ---
>>> I ran the Travis CI tests with no error reported:
>>> https://travis-ci.org/xypron2/u-boot/builds/274042004
>>>
>>> It is unclear to me why on Debian Stretch the .got and .got.plt sections
>>> are generated. The problem does not exist on Debian Jessie which uses
>>> binutils 2.25.
>>>
>>> I found this comment in the BFD code which describes under which
>>> circumstances the .got.plt section is not generated:
>>>
>>> https://sourceware.org/viewvc/src/bfd/elf64-x86-64.c?view=markup#l2921
>>>
>>> Don't allocate .got.plt section if there are no GOT nor PLT
>>> entries and there is no reference to _GLOBAL_OFFSET_TABLE_.

Can we have some binutils guys to look at this? Why does the behavior
get changes between versions?

>>> ---
>>> scripts/Makefile.spl | 3 ++-
>>> 1 file changed, 2 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/scripts/Makefile.spl b/scripts/Makefile.spl
>>> index dd8065d87d..3b2d79d562 100644
>>> --- a/scripts/Makefile.spl
>>> +++ b/scripts/Makefile.spl
>>> @@ -286,7 +286,8 @@ quiet_cmd_objcopy = OBJCOPY $@
>>> cmd_objcopy = $(OBJCOPY) $(OBJCOPYFLAGS) $(OBJCOPYFLAGS_$(@F)) $< $@
>>>
>>> OBJCOPYFLAGS_$(SPL_BIN)-nodtb.bin = $(SPL_OBJCFLAGS) -O binary \
>>> -            $(if $(CONFIG_SPL_X86_16BIT_INIT),-R .start16 -R .resetvec)
>>> +            $(if $(CONFIG_SPL_X86_16BIT_INIT), \
>>> +            -R .start16 -R .resetvec -R .got -R .got.plt)
>>

Simon, would you take a look at this?

>> If there is a non-empty .got, it can not be safely removed. The remaining code
>> can (and will) reference data within the .got.
>> Did you check the contents of the .got and where these are referenced from?
>
> This is from spl/u-boot-spl.map:
>
> .text           0x00000000fffd0000     0xa032
>  *(.text*)
>  .text          0x00000000fffd0000       0xcc arch/x86/cpu/start.o
>                 0x00000000fffd0000                _x86boot_start
>                 0x00000000fffd0020                _start
>
> .got            0x0000000000120000        0x4
>  .got           0x0000000000120000        0x4 arch/x86/cpu/start.o
>
> .got.plt        0x0000000000120004        0xc
>  .got.plt       0x0000000000120004        0xc arch/x86/cpu/start.o
>                 0x0000000000120004                _GLOBAL_OFFSET_TABLE_
>
> So in .got there is just the reference to the BIOS start address.
> And .got.plt contains additionally a pointer to .got.
>
>> Did you check the contents of the .got and where these are referenced
>> from?
> It is unclear to me what additional information is needed and how it can
> be extracted.
>

Regards,
Bin