diff --git a/Include/pystate.h b/Include/pystate.h
--- a/Include/pystate.h
+++ b/Include/pystate.h
@@ -8,6 +8,10 @@
 extern "C" {
 #endif
 
+#if defined(WITH_THREAD) && !defined(Py_LIMITED_API)
+#include "pythread.h"
+#endif
+
 /* State shared between threads */
 
 struct _ts; /* Forward */
@@ -118,6 +122,11 @@ typedef struct _ts {
     int trash_delete_nesting;
     PyObject *trash_delete_later;
 
+    /* Called when a thread state is deleted normally, but not when it
+       is destroyed after fork(). */
+    void (*on_delete)(void *);
+    void *on_delete_data;
+
     /* XXX signal handlers should also be here */
 
 } PyThreadState;
diff --git a/Lib/_dummy_thread.py b/Lib/_dummy_thread.py
--- a/Lib/_dummy_thread.py
+++ b/Lib/_dummy_thread.py
@@ -81,6 +81,10 @@ def stack_size(size=None):
         raise error("setting thread stack size not supported")
     return 0
 
+def _set_sentinel():
+    """Dummy implementation of _thread._set_sentinel()."""
+    return LockType()
+
 class LockType(object):
     """Class implementing dummy implementation of _thread.LockType.
 
diff --git a/Lib/test/test_threading.py b/Lib/test/test_threading.py
--- a/Lib/test/test_threading.py
+++ b/Lib/test/test_threading.py
@@ -444,6 +444,31 @@ class ThreadTests(BaseTestCase):
         self.assertEqual(out, b'')
         self.assertEqual(err, b'')
 
+    def test_tstate_lock(self):
+        # Test an implementation detail of Thread objects.
+        started = _thread.allocate_lock()
+        finish = _thread.allocate_lock()
+        started.acquire()
+        finish.acquire()
+        def f():
+            started.release()
+            finish.acquire()
+            time.sleep(0.01)
+        # The tstate lock is None until the thread is started
+        t = threading.Thread(target=f)
+        self.assertIs(t._tstate_lock, None)
+        t.start()
+        started.acquire()
+        # The tstate lock can't be acquired when the thread is running
+        # (or suspended).
+        tstate_lock = t._tstate_lock
+        self.assertFalse(tstate_lock.acquire(timeout=0), False)
+        finish.release()
+        # When the thread ends, the state_lock can be successfully
+        # acquired.
+        self.assertTrue(tstate_lock.acquire(timeout=5), False)
+        self.assertFalse(t.is_alive())
+
 
 class ThreadJoinOnShutdown(BaseTestCase):
 
diff --git a/Lib/threading.py b/Lib/threading.py
--- a/Lib/threading.py
+++ b/Lib/threading.py
@@ -33,6 +33,7 @@ except ImportError:
 # Rename some stuff so "from threading import *" is safe
 _start_new_thread = _thread.start_new_thread
 _allocate_lock = _thread.allocate_lock
+_set_sentinel = _thread._set_sentinel
 get_ident = _thread.get_ident
 ThreadError = _thread.error
 try:
@@ -548,6 +549,7 @@ class Thread:
         else:
             self._daemonic = current_thread().daemon
         self._ident = None
+        self._tstate_lock = None
         self._started = Event()
         self._stopped = False
         self._block = Condition(Lock())
@@ -625,9 +627,18 @@ class Thread:
     def _set_ident(self):
         self._ident = get_ident()
 
+    def _set_tstate_lock(self):
+        """
+        Set a lock object which will be released by the interpreter when
+        the underlying thread state (see pystate.h) gets deleted.
+        """
+        self._tstate_lock = _set_sentinel()
+        self._tstate_lock.acquire()
+
     def _bootstrap_inner(self):
         try:
             self._set_ident()
+            self._set_tstate_lock()
             self._started.set()
             with _active_limbo_lock:
                 _active[self._ident] = self
@@ -741,16 +752,17 @@ class Thread:
 
         self._block.acquire()
         try:
-            if timeout is None:
-                while not self._stopped:
-                    self._block.wait()
-            else:
-                deadline = _time() + timeout
-                while not self._stopped:
-                    delay = deadline - _time()
-                    if delay <= 0:
-                        break
-                    self._block.wait(delay)
+            def pred():
+                if self._tstate_lock is not None:
+                    # Ensure the underlying thread state has been
+                    # unregistered from the interpreter (see issue #18808).
+                    return self._stopped and self._tstate_lock.acquire()
+                else:
+                    # Main threads and dummy threads don't have a thread
+                    # state lock.
+                    return self._stopped
+            self._block.wait_for(pred, timeout)
+            self._tstate_lock = None
         finally:
             self._block.release()
 
diff --git a/Modules/_threadmodule.c b/Modules/_threadmodule.c
--- a/Modules/_threadmodule.c
+++ b/Modules/_threadmodule.c
@@ -1172,6 +1172,61 @@ yet finished.\n\
 This function is meant for internal and specialized purposes only.\n\
 In most applications `threading.enumerate()` should be used instead.");
 
+static void
+release_sentinel(void *wr)
+{
+    /* Tricky: this function is called when the current thread state
+       is being deleted.  Therefore, only simple C code can safely
+       execute here. */
+    PyObject *obj = PyWeakref_GET_OBJECT(wr);
+    lockobject *lock;
+    if (obj != Py_None) {
+        assert(Py_TYPE(obj) == &Locktype);
+        lock = (lockobject *) obj;
+        if (lock->locked) {
+            PyThread_release_lock(lock->lock_lock);
+            lock->locked = 0;
+        }
+    }
+    /* Deallocating a weakref with a NULL callback only calls
+       PyObject_GC_Del(), which can't call any Python code. */
+    Py_DECREF(wr);
+}
+
+static PyObject *
+thread__set_sentinel(PyObject *self)
+{
+    PyObject *wr;
+    PyThreadState *tstate = PyThreadState_Get();
+    lockobject *lock = newlockobject();
+
+    if (tstate->on_delete != NULL) {
+        PyErr_SetString(PyExc_RuntimeError,
+                        "Can't set sentinel for the current thread state");
+        return NULL;
+    }
+    if (lock == NULL)
+        return NULL;
+    /* The lock is owned by whoever called _set_sentinel(), but the weakref
+       hangs to the thread state. */
+    wr = PyWeakref_NewRef((PyObject *) lock, NULL);
+    if (wr == NULL) {
+        Py_DECREF(lock);
+        return NULL;
+    }
+    tstate->on_delete_data = (void *) wr;
+    tstate->on_delete = &release_sentinel;
+    return (PyObject *) lock;
+}
+
+PyDoc_STRVAR(_set_sentinel_doc,
+"_set_sentinel() -> lock\n\
+\n\
+Set a sentinel lock that will be released when the current thread\n\
+state is finalized (after it is untied from the interpreter).\n\
+\n\
+This is a private API for the threading module.");
+
 static PyObject *
 thread_stack_size(PyObject *self, PyObject *args)
 {
@@ -1247,6 +1302,8 @@ static PyMethodDef thread_methods[] = {
      METH_NOARGS, _count_doc},
     {"stack_size",              (PyCFunction)thread_stack_size,
      METH_VARARGS, stack_size_doc},
+    {"_set_sentinel",           (PyCFunction)thread__set_sentinel,
+     METH_NOARGS, _set_sentinel_doc},
     {NULL,                      NULL}           /* sentinel */
 };
 
diff --git a/Python/pystate.c b/Python/pystate.c
--- a/Python/pystate.c
+++ b/Python/pystate.c
@@ -208,6 +208,8 @@ new_threadstate(PyInterpreterState *inte
 
         tstate->trash_delete_nesting = 0;
         tstate->trash_delete_later = NULL;
+        tstate->on_delete = NULL;
+        tstate->on_delete_data = NULL;
 
         if (init)
             _PyThreadState_Init(tstate);
@@ -390,6 +392,9 @@ tstate_delete_common(PyThreadState *tsta
     if (tstate->next)
         tstate->next->prev = tstate->prev;
     HEAD_UNLOCK();
+    if (tstate->on_delete != NULL) {
+        tstate->on_delete(tstate->on_delete_data);
+    }
     PyMem_RawFree(tstate);
 }