google-protobuf won't build on Linux after adding MinGW platform

[AnyOldName3]

I had a working Jekyll project that built fine under both real Linux and WSL2, but for reasons not worth getting into, needed to switch from WSL2 to Windows. I ran bundle lock --add-platform x64-mingw-ucrt01, the results of which can be seen as this commit: openscenegraph/openscenegraph.io@a72328e. The line I believe to be most important changes google-protobuf (3.25.1-x86_64-linux) to google-protobuf (3.25.1), and looking at the contents of the two packages, it appears the former has prebuilt .sos whereas the latter is source-only, so the shared library needs building.

As intended, this made the project work with Windows Ruby, however, it also produced the following errors on Linux:

Gem::Ext::BuildError: ERROR: Failed to build gem native extension.

    current directory: /mnt/c/vsg/osg-website-replacement/vendor/bundler/ruby/3.0.0/gems/google-protobuf-3.25.1/ext/google/protobuf_c
/usr/bin/ruby3.0 -I /usr/lib/ruby/vendor_ruby -r ./siteconf20231220-407-i7czk0.rb extconf.rb
creating Makefile

current directory: /mnt/c/vsg/osg-website-replacement/vendor/bundler/ruby/3.0.0/gems/google-protobuf-3.25.1/ext/google/protobuf_c
make DESTDIR\= clean

current directory: /mnt/c/vsg/osg-website-replacement/vendor/bundler/ruby/3.0.0/gems/google-protobuf-3.25.1/ext/google/protobuf_c
make DESTDIR\=
compiling convert.c
convert.c: In function ‘Msgval_IsEqual’:
convert.c:295:58: warning: passing argument 4 of ‘shared_Msgval_IsEqual’ discards ‘const’ qualifier from pointer target type [-Wdiscarded-qualifiers]
  295 |                                             type_info.def.msgdef, &status);
      |                                             ~~~~~~~~~~~~~^~~~~~~
In file included from convert.c:21:
shared_convert.h:19:60: note: expected ‘upb_MessageDef *’ but argument is of type ‘const upb_MessageDef *’
   19 |                            upb_CType type, upb_MessageDef* msgdef,
      |                                            ~~~~~~~~~~~~~~~~^~~~~~
convert.c:299:5: error: format not a string literal and no format arguments [-Werror=format-security]
  299 |     rb_raise(rb_eRuntimeError, upb_Status_ErrorMessage(&status));
      |     ^~~~~~~~
convert.c: In function ‘Msgval_GetHash’:
convert.c:308:41: warning: passing argument 3 of ‘shared_Msgval_GetHash’ discards ‘const’ qualifier from pointer target type [-Wdiscarded-qualifiers]
  308 |       val, type_info.type, type_info.def.msgdef, seed, &status);
      |                            ~~~~~~~~~~~~~^~~~~~~
In file included from convert.c:21:
shared_convert.h:23:48: note: expected ‘upb_MessageDef *’ but argument is of type ‘const upb_MessageDef *’
   23 |                                upb_MessageDef* msgdef, uint64_t seed,
      |                                ~~~~~~~~~~~~~~~~^~~~~~
convert.c:312:5: error: format not a string literal and no format arguments [-Werror=format-security]
  312 |     rb_raise(rb_eRuntimeError, upb_Status_ErrorMessage(&status));
      |     ^~~~~~~~
cc1: some warnings being treated as errors
make: *** [Makefile:245: convert.o] Error 1

make failed, exit code 2

This log is from using WSL2, but it also happens in CI: https://github.com/AnyOldName3/osg-website-replacement/actions/runs/7266847247/job/19829238652. I believe that's Arch running in Docker (on an Ubuntu host) whereas WSL2 is running Ubuntu, so dependency packages from the OS should be different - they're at least using different versions of Ruby.

-Werror=format-security is coming from rbconfig.rb on the WSL2 system, and I imagine that's where it's coming from on the probably-Arch VM, too. Conceivably, other compiler flags requested by that file could be responsible for the other errors. I don't know whether this is a case of a broken gem (i.e. that these flags should work), a packaging error for Ruby itself (i.e. that these flags shouldn't have ended up in rbconfig.rb, or some other kind of problem. I also don't know how I'd go about working around this, e.g. by somehow forcing Gemfile.lock to use the prebuilt Linux variant of the package on Linux.

[indirect]

This sounds like a bug to me, it should be possible to keep the platform-specific version of the gem locked when running on the relevant platform. The only workaround that comes to mind for now is having a separate Gemfile.lock per platform.

[AnyOldName3]

Whose bug is it? Do I need to report it on another project's tracker?

[indirect]

Let's check with @deivid-rodriguez and see if we should open this as a bug against Bundler

[deivid-rodriguez]

Yes, that looks like a bug to me. There's no reason to remove the Linux platform specific version. Let me move this to an issue.

[AnyOldName3]

To make it easier to go back and forth, I'll link the issue: #7321