[RFC 4/4] ld: bfd: sframe: fix incorrect r_addend in RELA entries

[Jose E. Marchesi]

Hi Jan.

> On 10.03.2025 14:16, Jens Remus wrote:
>> On 08.03.2025 08:38, Indu Bhagat wrote:
>>> With the fix in GAS, we now use a different PC-relative RELA for
>>> updating the SFrame FDE function start address: The value is the offset
>>> of the start PC of the function from the start of the SFrame section.
>>>
>>> When RELAs are output (e.g. for relocatable links), there is need to
>>> adjust the r_addend.  This is because the r_addend values still have the
>>> correct values for the _input_ SFrame section being linked / relocated.
>>> The values must now be (before outputing the RELAs) with respect to the
>>> _output_ SFrame section.
>>>
>>> PS: This patch should be merged with the previous commits before final
>>> commit (Otherwise the tests will fail).  It is currently a separate
>>> patch as I would like to check if this is OK to do.  If such a "addend
>>> fixup" is risky or wrong, it seems we will need a new type of RELOC for
>>> SFrame sections.
>> 
>>> diff --git a/bfd/elf-sframe.c b/bfd/elf-sframe.c
>> 
>>> @@ -574,6 +574,37 @@ _bfd_elf_sframe_section_offset (bfd *output_bfd ATTRIBUTE_UNUSED,
>>>   	  + out_fde_idx * sizeof (sframe_func_desc_entry));
>>>   }
>>>   
>>> +/* Get the "canonicalized" addend for the symbol reference corresponding to the
>>> +   relocation at RELOC_INDEX.  E.g.,  for the following reloc for the SFrame
>> 
>> Nit: s/reloc/relocs/
>> 
>>> +   FDE function start address:
>>> +	Offset          Type          Sym. Name + Addend
>>> +	00000000001c  R_X86_64_PC32   .text + 1c
>>> +	000000000030  R_X86_64_PC32   .text + 3b
>>> +   The canonicalized addend are 0 and b respectively as the relocs are for
>>> +   symbols (.text + 0) and (.text + b) respectively.
>> 
>> Maybe it would help to use the function symbols foo and bar in the example:
>> 
>>     relocation at RELOC_INDEX.  E.g.,  for the following annotated relocs for the SFrame
>>     FDE function start address of FDE[0] and FDE[1] for functions foo and bar:
>> 	Offset          Type          Sym. Name + Addend
>> 	00000000001c  R_X86_64_PC32   .text + 1c   // FDE[0] for foo at .text + 0
>> 	000000000030  R_X86_64_PC32   .text + 3b   // FDE[1] for bar at .text + b
>>     With:
>>       1c = 0 + sizeof(sframe_header) + 0 * sizeof(sframe_fde)
>>       3b = b + sizeof(sframe_header) + 1 * sizeof(sframe_fde)
>>     The canonicalized addend are 0 and b respectively as the relocs are for
>>     symbols (.text + 0) and (.text + b) respectively.
>
> Hmm, wait - there's a mix between .text and .sframe then, isn't there? That
> would explain why custom handling is then necessary. Yet as said in the other
> mail, custom handling should not be necessary. Not the least because other
> ELF-consuming tools also need to work, not just what GNU binutils provides.

[At a risk of being reiterative, but this also helps to my own
 understanding, because this is all very confusing :)]

I think there is nothing special about that "mix".  As far as I can tell
this is just one rather peculiar particular case of a PCrel reloc.
Suppose you have the typical:

  .text
  ...
    ldinsn OFFSET

  .data
  ...
  somedata:

And then you install a relocation for the instruction's offset, which is
PC-relative.  You use an expression like:

  ldinsn somedata

Then I would expect GAS to emit a PCrel with:

  .text.rela:
    P: offset_of_ldins_within_text
    S: .data
    A: offset_of_somedata_within_data

With resolving expression:

  S + A - P

to which the linker also adds .text, resulting in the expected:

  .data + offset_of_somedata_within_data - .text + offset_of_ldins_within_text

i.e. the distance between somedata and the instruction which field is to
be patched.

Now, for SFrame, the expression used wants to denote the distance
between the beginning of some given function and the beginning of the
.sframe section containing the FDE to patch.  GAS is given the
expression:

   .text + text_offset - .sframe

And the assembler smartly uses a PCrel relocation (it has no other
choice I suppose) such as:

  .sframe.rela:
    P: offset_of_fde_within_sframe
    S: .text
    A: offset_of_function_within_text + offset_of_fde_within_sframe
                                        ~~~~~~~~~~~~~~~~~~~~~~~~~~~

[I have been trying to find the code in GAS that actually does this,
 because it seems to me it must be some ad-hoc code, but couln't find
 it.]

Which is a little weird, because it has the additional
offset_of_fde_within_sframe, intended to basically annull the same value
in P in the resolving function:

   S + A - P

which is, with the linker-added .sframe:

   .text + (offset_of_funtion_within_text + offset_of_fde_within_sframe)
         - .sframe + offset_of_fde_within_sframe

which becomes:

   .text + offset_of_funtion_within_text - .sframe

which is the desired distance between the beginning of the function and
the beginning of the .sframe section.

I don't think ELF consuming tools that just apply relocations need any
special code to solve these relocs.

But due how the SFrame format works, as Jens explained, tools that merge
SFrame input sections into a relocatable output section need to
basically recreate the relocations for the FDEs in the output.  If this
merging consisted (like with ehframe) just in collating the input
sections together, then no special casing would be needed.  But merging
SFrame sections imply that the offset of the FDEs wrt the containing
output .sframe will likely be different.  This offset is embedded in the
input relocation's offset and addend, as shown above.  So the merging
tool has no other choice that to "remove" the old FDE offset from the
input relocation (from both addend and offset) and "insert" the new FDE
offset (in the resulting merged .sframe section in the output) in both
addend and offset.  Tools that merge SFrame sections must use
sframe-specific logic anyway, so I don't think this is a big deal...

The other changes in generic code in Indu's patch, as far as I can see,
are necessary because the linker assumes that the input section contents
are merged by stocking them (or a subset of them) in the merged output
section.  This is not so in SFrame, so thats why Indu's patch makes ld
to not increase that offset within the output section every iteration in
the loop, if the sections being merged are .sframe.  It is simply that
until now no format required this sort of merging, where I11,I12,I13 and
I21,I22,I23 areas in two input sections will be merged into for example
I11,I21,I12,I22,I13,I23 in the output section.