diff -r acd31653e3b4 Doc/c-api/gcsupport.rst
--- a/Doc/c-api/gcsupport.rst	Thu Jan 16 06:53:54 2014 +0100
+++ b/Doc/c-api/gcsupport.rst	Thu Jan 16 14:48:37 2014 +0100
@@ -150,3 +150,34 @@
    this method (don't just call :c:func:`Py_DECREF` on a reference).  The
    collector will call this method if it detects that this object is involved
    in a reference cycle.
+
+Practical Considerations
+------------------------
+
+From the perspective of an extension writer, correctly implementing support for
+cyclic garbage collection is easier than it appears from the above description.
+If no custom :c:member:`~PyTypeObject.tp_alloc` and
+:c:member:`~PyTypeObject.tp_free` functions are specified, the default handlers
+will adhere to the GC rules for constructors and deallocators.
+In that case, specifying the :const:`Py_TPFLAGS_HAVE_GC` flag, and implementing
+the :c:member:`~PyTypeObject.tp_traverse` and
+:c:member:`~PyTypeObject.tp_clear` handlers is enough to fully support cyclic
+garbage collection. Most extension writers will never need to worry about
+:c:func:`PyObject_GC_Track` and :c:func:`PyObject_GC_UnTrack` or how memory is
+allocated.
+
+In contrast, container types supporting cyclic garbage collection while
+implementing custom memory management require special thought.
+The rule for memory to be allocated with :c:func:`PyObject_GC_New` or
+:c:func:`PyObject_GC_NewVar` prevents the use of other memory allocation
+backends. Placing several Python objects in a larger memory pool previously
+allocated via :c:func:`PyObject_GC_New` is not supported either. Consequently,
+custom allocators can in practice only rely on :c:func:`PyObject_GC_New` or
+:c:func:`PyObject_GC_NewVar`, while taking care to call
+:c:func:`PyObject_GC_Track` before returning.
+
+If truly custom memory allocation is required for extension classes, the
+:const:`Py_TPFLAGS_HAVE_GC` flag must be omitted. The resulting objects will
+still be destroyed as long as they are not part of any reference cycle.
+Possible ways to meet this requirement are to not hold references to other
+Python objects or to judiciously use weak references.