I think that we still need to restore errno because it will be called just after mmap is created for builin case

_PyAnnotateMemoryMap() is only called when a memory allocation succeed. Would you mind to elaborate in which case it matters to preserve the previous errno?

mmap called -> errno 0 -> _PyAnnotateMemoryMap called -> errno non 0 -> propagate errno?

Since _PyAnnotateMemoryMap is optional, I would like to avoid this errno is exposed to caller.

In Modules/mmapmodule.c, errno is only used if mmap() fails:

    Py_BEGIN_ALLOW_THREADS
    m_obj->data = mmap(NULL, map_size, prot, flags, fd, offset);
    Py_END_ALLOW_THREADS

    int saved_errno = errno;
    if (devzero != -1) {
        close(devzero);
    }

    if (m_obj->data == (char *)-1) {
        m_obj->data = NULL;
        Py_DECREF(m_obj);
        errno = saved_errno;
        PyErr_SetFromErrno(PyExc_OSError);
        return NULL;
    }
#ifdef MAP_ANONYMOUS
    if (m_obj->flags & MAP_ANONYMOUS) {
        (void)_PyAnnotateMemoryMap(m_obj->data, map_size, "cpython:mmap");
    }
#endif
    m_obj->access = (access_mode)access;
    return (PyObject *)m_obj;

I don't see what you mean by propagate errno, since errno is not used when _PyAnnotateMemoryMap() is called.

Get it!