GeezerVison

Following functions will be moved here in the end.

• _PyTuple_ClearFreeList
• _PyFloat_ClearFreeList
• _PyDict_ClearFreeList
• _PyAsyncGen_ClearFreeLists
• _PyContext_ClearFreeList

Following functions will be moved here in the end.

• _PyDict_Fini
• _PyTuple_Fini
• _PySlice_Fini
• _PyFloat_Fini
• _PyContext_Fini

Hummm, I find this asymmetry a bit odd. Wouldn't be better to keep all functions receiving PyInterpreterState *interp and then fetching the freelist state from there instead?

That way the code of the caller will be also cleaner and doesn't need to deal with where the freelist is stored.

In the free-threaded builds, we want to clear the free-lists for a specific thread (not all threads). Passing interp doesn't provide enough information. In some cases the thread we are clearing is not even the active thread, such as when we call PyThreadState_Clear() on dead threads after fork.

This locks interpreters creation, but not thread creation, no? Is it possible that this races with thread creation/destruction? I am missing something?

This part looks okay to me. The HEAD_LOCK() protects the threads linked list (among other things). Threads acquire the HEAD_LOCK() when appending (or removing) themselves to that list.

Modification of the free-lists themselves is protected by the "stop-the-world" pause (or the GIL), which is not yet implemented.

This part looks okay to me. The HEAD_LOCK() protects the threads linked list (among other things). Threads acquire the HEAD_LOCK() when appending (or removing) themselves to that list.

Modification of the free-lists themselves is protected by the "stop-the-world" pause (or the GIL), which is not yet implemented.

Can we pass something like int is_finalization as an argument to _Py_ClearFreeLists()? I am concerned about the issue that _PyList_Fini() attempts to detect by setting num_free=-1: allocating objects after the free list is cleared for the final time. We want that error detection here.

Something like:

_Py_ClearFreeLists(freelist_state, 1);

and in the code clearing free lists

_PyList_ClearFreeList(_PyFreeListState *freelist_state, int is_finalization)
{
    ....
   if (is_finalization) {
     state->numfree = -1;
   }
}

Along those lines, I think we should move the _Py_ClearFreeLists() call after the on_delete() call since, on_delete() can free some Python objects.

@colesbury We may need to change the assertion of

If we mark it at the PyThreadState_Clear phase.

In the free-threaded builds, we want to clear the free-lists for a specific thread (not all threads). Passing interp doesn't provide enough information. In some cases the thread we are clearing is not even the active thread, such as when we call PyThreadState_Clear() on dead threads after fork.