mirror of
https://github.com/nushell/nushell.git
synced 2026-11-22 17:27:05 +08:00
## Description On unix, importing a module that imports one of its parents back through a `..` path never terminates and nushell dies with a segfault instead of reporting a circular import (#18731): ```nu mkdir /tmp/nucrash/c/i "export module i" | save /tmp/nucrash/c/mod.nu "use ../mod.nu" | save /tmp/nucrash/c/i/mod.nu nu -c "use /tmp/nucrash/c" # Segmentation fault (core dumped) ``` The parser builds these paths with `std::path::absolute`, which drops `.` components but keeps `..` on unix, because a `..` component is not the same thing as the parent directory once symlinks are in play. `FileStack`, which is what detects circular imports, compared those paths as they were, so `/tmp/nucrash/c/mod.nu` and `/tmp/nucrash/c/i/../mod.nu` looked like two different files. Each trip around the cycle appended another `/i/..` to the path, so the import never repeated itself, the check never fired, and the parser recursed until the stack ran out. That also explains why the symptom is not the same everywhere. Whether you get a crash depends on which limit the growing path hits first: on macOS the path stops resolving before the stack runs out, so it surfaces as a `module_not_found` naming a path with several hundred `../i` components in it, as @pickx saw on the issue. Windows was already fine, because `std::path::absolute` resolves `..` there, so the same repro reports `nu::parser::circular_import` today. This makes the other platforms behave that way too. Files on the stack are now identified by their canonical path, which resolves `..` along with symlinks pointing at a file that is already being parsed. The path a file was actually reached by is kept as it was, since that is what error messages show and what the files it imports are resolved against, so nothing else changes. Paths with no file behind them, such as the virtual paths of the standard library, fall back to expanding `.` and `..` textually. Legitimate `..` imports are unaffected: a module importing a sibling through `../other/mod.nu` still resolves and is not reported as circular (`module_nested_imports_in_dirs` covers this). Tested on Linux, where the repro now prints a circular import error and exits 1 instead of dying on SIGSEGV, and on Windows, where behavior is unchanged. The new regression test fails on Linux without the fix and passes with it. ## User-facing changes (Release notes) Fixed a crash when importing a module that imports one of its parents back through a `..` path, for example a `c/i/mod.nu` containing `use ../mod.nu`. Nushell reported a circular import on Windows but crashed on other platforms, and now reports the circular import everywhere. ## Additional notes fixes #18731