Fix bug in optimizing module loader for coverage collection
Summary: In coverage collection mode a special module loader is prepended to `sys.meta_path`. In very specific conditions this module loader can end up returning a loader pointing to a _completely wrong module_. When importing symbols from the wrong module errors occur. The conditions to trigger the bug are: - running in coverage collection mode, enabling the custom loader - the test binary is a zip (e.g. par_style=fastzip) - having a module name where the end part matches the name of a builtin Python module When these conditions were met, the special loader would return the builtin Python module instead of the expected module. E.g. when loading a module like `myapp.somemod.platform` in a zip style binary. The custom loader first calls `imp.find_module()` to find the module it wants to return a wrapped loader for. This fails for modules included in the test binary, because the builtin `imp` module can not load from zips. This was the trigger leading to the call to the buggy code. When the initial call to `imp.find_module()` failed, the custom loader would try a second call, asking the internal loader to this time try any path on `sys.path`. For most module names this call would also fail, making the custom loader return `None`, after which Python tries other loaders on `sys.path`. However, when the final part of the module that was asked to load matches the name of a Python builtin module, then the second call to the `imp` module would succeed, returning a loader for the builtin module. E.g. `platform` when asking for `myapp.somemod.platform`. This diff fixes the issue by removing the broken second call to the internal loader. This will never have worked, we just have not triggered or noticed triggering the wrong loading before. Differential Revision: D20798119 fbshipit-source-id: dffb54e308106a81af21b63c5ee64c6ca2041920
Showing
Please register or sign in to comment