rdata is supposed to be read-only isn't it?

Due to a linker bug mutable globals are ending up here. I don't understand this enough to open an issue, but @kubkon said to just do this for now.

For reference, the global I ran into this with is lib.test_runner.cmdline_buffer, where the test program would crash on first write to this buffer, and only when the emitted code reaches a certain size (such as enabling n tests, or making unrelated changes to what code is emitted).

It's not rdata, it's .data.rel.ro since we may have pointers inside a struct we lower in that section. The linker code path responsible for selecting output section doesn't do a recursive check for embedded pointers in any linker (macho, elf or coff), and this is important for PIC/PIE targets where the dynamic linker has to slide the address of each pointer within the binary - hence .data.rel.ro which is only writeable during dynamic relocation startup phase - well, in theory at least.

Anyhow, as @jacobly0 mentioned, that was my suggestion to make it writeable to unblock his progress. I am aware of this nuance and will provide a fix in ELF first, and then backport it to other linkers as I start working on them.