在 2023/5/22 05:13, Mark Wielaard 写道:
> Hi Ying,
>
> On Tue, May 16, 2023 at 02:38:45PM +0800, Ying Huang wrote:
>> 在 2023/5/9 23:15, Mark Wielaard 写道:
>>> On Tue, 2023-04-11 at 16:12 +0800, Ying Huang wrote:
>>>> In mips64 little-endian, r_info consists of four byte fields(contains
>>>> three reloc types) and a 32-bit symbol index. In order to adapt
>>>> GELF_R_SYM and GELF_R_TYPE, need convert raw data to get correct symbol
>>>> index and type.
>>> Is there a spec that describes this?
>>    
>>
>>  references:
>>
>>     https://www.linux-mips.org/pub/linux/mips/doc/ABI/elf64-2.4.pdf
>>
>>     Page40 && Page41
> Thanks. Interesting. If possible please include this URL in a comment.
    OK.
>
>>> I see you adjusted elf.h to include:
>>>
>>> +#define ELF64_MIPS_R_TYPE(i)		((i) & 0xff)
>>> +#define ELF64_MIPS_R_TYPE2(i)           (((i) >> 8) & 0xff)
>>> +#define ELF64_MIPS_R_TYPE3(i)           (((i) >> 16) & 0xff)
>>>
>>> And various new relocation types for MIPS variants.
>>> Our elf.h comes from the glibc project. Have you also submitted these
>>> changes to libc-alpha@sourceware.org ?
>>     Has submitted.
>>
>>     https://sourceware.org/pipermail/libc-alpha/2023-May/148026.html
>>
>>     https://sourceware.org/pipermail/libc-alpha/2023-May/148112.html
> Thanks. We'll sync them as soon as v3 lands.
>
>>     Has tested on big/little endian machine with big/little endian
>> ELF file, all were OK.
> Could you post those ELF files somewhere?
> Maybe we can add some as test files?
    The ELF files I used were "src/strip.o" which were generated on big/little endian machine.

root@debian-sid-mipsbe:~# readelf -h strip.o
ELF Header:
  Magic:   7f 45 4c 46 02 01 01 00 00 00 00 00 00 00 00 00
  Class:                             ELF64
  Data:                              2's complement, little endian
  Version:                           1 (current)
  OS/ABI:                            UNIX - System V
  ABI Version:                       0
  Type:                              REL (Relocatable file)
  Machine:                           MIPS R3000
  Version:                           0x1
  Entry point address:               0x0
  Start of program headers:          0 (bytes into file)
  Start of section headers:          426696 (bytes into file)
  Flags:                             0x80000007, noreorder, pic, cpic, mips64r2
  Size of this header:               64 (bytes)
  Size of program headers:           0 (bytes)
  Number of program headers:         0
  Size of section headers:           64 (bytes)
  Number of section headers:         41
  Section header string table index: 40
root@debian-sid-mipsbe:~# readelf -h elfutils_debug/src/strip.o
ELF Header:
  Magic:   7f 45 4c 46 01 02 01 00 00 00 00 00 00 00 00 00
  Class:                             ELF32
  Data:                              2's complement, big endian
  Version:                           1 (current)
  OS/ABI:                            UNIX - System V
  ABI Version:                       0
  Type:                              REL (Relocatable file)
  Machine:                           MIPS R3000
  Version:                           0x1
  Entry point address:               0x0
  Start of program headers:          0 (bytes into file)
  Start of section headers:          215248 (bytes into file)
  Flags:                             0x70001007, noreorder, pic, cpic, o32, mips32r2
  Size of this header:               52 (bytes)
  Size of program headers:           0 (bytes)
  Number of program headers:         0
  Size of section headers:           40 (bytes)
  Number of section headers:         40
  Section header string table index: 39
root@debian-sid-mipsbe:~#

>>> It feels like this is in the wrong place and doesn't take into account
>>> whether the program itself is running on a big or little endian
>>> machine.
>>>
>>> Maybe I am misunderstanding why this conversion is needed always. But I
>>> would have expected a specific conversion function for MIPS for
>>> ELF_T_REL and/or ELF_T_RELA (which are currently generic).
>>     1.Considering that some people directly use the original data instead of the interface function gelf_getrela;
>>
>>     2.Because there is only one parameter, so can not modify the macro ELF64_R_TYPE；
>>
>>     3.If the mips interface function is added, other packages using libelf will not be compatible with mips, and the place where gelf_getrela is used need be modified.
>>
>>     Where do you think is the most suitable place to do mips conversion?
> So for the convert_data case in elf_getdata.c I would expect a special
> conversion function being defined for ELF_T_REL and ELF_T_RELA. This
> normally comes from __elf_xfctstom. I realize now that cannot easily
> take an Elf of machine type. But I would expect to have the conversion
> be done by a new function that can be assigned to fp in convert_data:
>
>       /* Get the conversion function.  */
>       fp = __elf_xfctstom[eclass - 1][type];
        Maybe I can create a new function named "convert_data_for_mips" and contained the code that added in convert_data?
>
> I still don't fully understand the need of the code in elf_update. I
> assume that is needed in case the file was used with ELF_C_RDWR_MMAP?
>
> Cheers,
>
> Mark

    Because we convert the relocation data for mips64el in elf_getdata.c, eg src/strip, and when we write elf  to file, need confirm the relocation data is in original order bits.

    When file use ELF_C_RDWR_MMAP, only impact whether malloc data_base in convert_data.


    Thanks,

    Ying