invalid relocation for non-allocated sections: R_X86_64_GOTPC64

[faern]

I'm trying to switch over to mold for linking our Rust projects at https://github.com/mullvad/mullvadvpn-app/. But I get the following error when building Rust binaries that pull in our static library archive for libnftnl:

  = note: mold: fatal: /home/faern/src/mullvadvpn-app/dist-assets/binaries/x86_64-unknown-linux-gnu/libnftnl.a(utils.o):(.debug_info): invalid relocation for non-allocated sections: R_X86_64_GOTPC64
          clang-14: error: linker command failed with exit code 1 (use -v to see invocation)

I can avoid the error and link fine if I rebuild libnftnl.a with either gcc or clang directly on my Fedora development machine (gcc 12.2.1 and clang 14.0.5) or with clang in our Debian 10 based build container (clang version 7.0.1-8+deb10u2). However, building libnftnl.a with gcc in the Debian 10 build container (gcc version 8.3.0)) is what triggers the error above.

Is the only recommended solution simply to use clang / upgrade the build environment, or is this issue fixable in mold in some way that allows us to continue building with gcc on Debian 10? I would prefer to not have to change compiler just for this, if possible.

Here is the Makefile used for building libnftnl. And here is the container definition.

Thank you for an amazingly fast linker <3

[rui314]

Thank you for your report! This should be fixed in mold. I'll make a change to support that relocation type.

[rui314]

On second thought, it is odd that a debug section contains a PC-relative relocation because the notion of PC-relative doesn't make sense for non-memory-mapped section. Can you upload your /home/faern/src/mullvadvpn-app/dist-assets/binaries/x86_64-unknown-linux-gnu/libnftnl.a so that I can take a look into it? It is likely a compiler bug that happens to work with GNU ld.

[faern]

Absolutely. The version checked in to our repository has this issue. It's very old and built in Debian 9, but gives the same error. It's available here.

But I'll also upload a freshly built one that has the same issue but built on a newer build container (Debian 10). For comparison I'll also upload this one built in the same container from the exact same Makefile and source code, but with clang:

• libnftnl.a-gcc-broken.zip.
• libnftnl.a-clang-working.zip

[rui314]

Can you also run the final link command with -Wl,--repro? If that flag is given, mold creates a tar file containing all input files with which it is very easy to re-run the same command with the same output. Please share that file with me.

[faern]

Github will only allow me to upload 25 MiB files, and this repro tar achive is 1.3 GiB (~150MiB when compressed).... And I can't push it to a temporary git branch either because there files can only be 100 MiB... I'll sort this out somehow.

[faern]

Here we go! The extra layer of zip is just because that's the only format Github allowed 😅

mullvad_daemon-96dd93fc4006358c.repro.tar.xz.zip

EDIT: I should probably also mention my mold version:

$ mold --version
mold 1.10.0 (compatible with GNU ld)

[rui314]

This is apparently a bug of an old version of GCC. According to its .comment section, utils.o was built by GCC 6.3.0. Are you sure you are using that old compiler? It was released in 2016.

[faern]

The build I uploaded the repro from is indeed very old. We have not had any reason to rebuild libnftnl for a couple of years since we have had no need to upgrade it. But if we were to rebuild it today we would use our Debian 10 build images (the oldest distro our software supports, so we want to link to the oldest supported glibc), and Debian 10 currently has gcc 8.3.0 and there I see the same error message.

I can create a new repro build with the 8.3.0 build if you think that makes a difference. But it would be awesome if mold would support the newest GCC supported by Debian 10 (it's still supported, albeit in LTS now)

[rui314]

Does the above change work for you?

[faern]

Interesting. This actually solves it for that specific version of libnftnl.a. However, when I re-build it with gcc 8.3.0 in Debian 10 I hit the same error again.. Was the fix too specific to 6.3.0?

Here is my repro for the failing build using libnftnl.a built by gcc 8.3.0 and with mold built from commit 77c4795:

mullvad_daemon-4eecad6dfc3f04fb.repro-gcc8.3.0.tar.xz

Thank you so much for paying attention to this gcc issue :)

[rui314]

Can you paste the new error message? It should be different from the previous one.

[faern]

  = note: mold: fatal: /home/faern/src/mullvadvpn-app/dist-assets/binaries/x86_64-unknown-linux-gnu/libnftnl.a(utils.o):(.debug_info): invalid relocation for non-allocated sections: R_X86_64_GOTOFF64
          clang-14: error: linker command failed with exit code 1 (use -v to see invocation)

I should not have trusted my eyes. I thought it looked the same, but it is indeed a different one!

[rui314]

Should be fixed in the above change.

[faern]

This works! Thank you so much!