[Default] Unsupported ELF relocation: R_X86_64_GOTOFF64

[xusheng6]

While analyzing this object file, we see the following error messages:

[Default] Unsupported ELF relocation: R_X86_64_GOTOFF64
[Default] Unsupported ELF relocation: R_X86_64_UNKNOWN27
[Default] Unsupported ELF relocation: R_X86_64_UNKNOWN31

As a result, we are not handling the call to _cpuid correctly:

liblzma_la-crc64-fast.o.zip

[ccarpenter04]

[Default warn] Unsupported ELF relocation: R_X86_64_DTPMOD64
[Default warn] Unsupported ELF relocation: R_X86_64_DTPOFF64

These relocation types are also unsupported

[fuzyll]

I tried looking at this tonight, and I don't think the issue with _cpuid is relevant to what's been identified in this ticket. As far as I can tell, it's R_X86_64_GOTPCREL for that call, which we've already implemented:

$ objdump -dr -j .text._get_cpuia liblzma_la-crc64-fast.o
...
  c7:   ff 15 00 00 00 00       call   *0x0(%rip)        # cd <.Llzma_block_param_encoder.0+0x59>
                        c9: R_X86_64_GOTPCREL   _cpuid-0x4
...

Also, since this is a .o file, I believe the output you're seeing is 100% expected? Aren't GOT entries in a .o placeholders until the linker fills them in later? I don't think there's anything we can do here to improve this whatsoever. But, if I'm wrong, someone please help me out!

In any case, I've implemented the missing relocations from the top of this issue and things look correct? But, I am unsure if they work outside of a .o file. I have not attempted the ones in the second comment because I don't have a test case to ensure they actually work.

Are either of you able to find these in other files so I can better test what I've done?

[xusheng6]

I tried looking at this tonight, and I don't think the issue with _cpuid is relevant to what's been identified in this ticket. As far as I can tell, it's R_X86_64_GOTPCREL for that call, which we've already implemented:

$ objdump -dr -j .text._get_cpuia liblzma_la-crc64-fast.o
...
  c7:   ff 15 00 00 00 00       call   *0x0(%rip)        # cd <.Llzma_block_param_encoder.0+0x59>
                        c9: R_X86_64_GOTPCREL   _cpuid-0x4
...

Also, since this is a .o file, I believe the output you're seeing is 100% expected? Aren't GOT entries in a .o placeholders until the linker fills them in later? I don't think there's anything we can do here to improve this whatsoever. But, if I'm wrong, someone please help me out!

In any case, I've implemented the missing relocations from the top of this issue and things look correct? But, I am unsure if they work outside of a .o file. I have not attempted the ones in the second comment because I don't have a test case to ensure they actually work.

Are either of you able to find these in other files so I can better test what I've done?

Hi, @fuzyll thanks for looking into this. I created this issue on behalf of a user. I am not super familiar with it, but I remember the context is that he tried to repro the analysis of the xz util backdoor in binja. And he runs into this call to cpuid

Now looking at it, I think you might be correct -- though ideally we can avoid generating an error when indeed we are already doing the right thing

[fuzyll]

I've implemented multiple relocations, including R_X86_64_GOTOFF64, in 19ea371. This should cover the things that were missing from the sample, even if it doesn't fix the _cpuid thing that was initially reported.

This is how things used to look, before builds >= 5.3.8955-dev:

...and this is how things look after:

I am leaving this issue open to track the relocations in the comment that @ccarpenter04 made and updating the title. I am also still hoping some user will come by and provide me a fully compiled and linked example with some of these relocations so I can make sure my code fully works. 🤞🏼