diff --git a/Include/pyport.h b/Include/pyport.h
index 62db8d07701d1d2..e77b39026a59c1f 100644
--- a/Include/pyport.h
+++ b/Include/pyport.h
@@ -509,23 +509,18 @@ extern "C" {
 #endif
 
 #ifdef WITH_THREAD
-#  ifdef Py_BUILD_CORE
-#    ifdef HAVE_THREAD_LOCAL
-#      error "HAVE_THREAD_LOCAL is already defined"
-#    endif
-#    define HAVE_THREAD_LOCAL 1
-#    ifdef thread_local
-#      define _Py_thread_local thread_local
-#    elif __STDC_VERSION__ >= 201112L && !defined(__STDC_NO_THREADS__)
-#      define _Py_thread_local _Thread_local
-#    elif defined(_MSC_VER)  /* AKA NT_THREADS */
-#      define _Py_thread_local __declspec(thread)
-#    elif defined(__GNUC__)  /* includes clang */
-#      define _Py_thread_local __thread
-#    else
-       // fall back to the PyThread_tss_*() API, or ignore.
-#      undef HAVE_THREAD_LOCAL
-#    endif
+// HAVE_THREAD_LOCAL is just defined here for compatibility's sake
+#  define HAVE_THREAD_LOCAL 1
+#  ifdef thread_local
+#    define _Py_thread_local thread_local
+#  elif __STDC_VERSION__ >= 201112L && !defined(__STDC_NO_THREADS__)
+#    define _Py_thread_local _Thread_local
+#  elif defined(_MSC_VER)  /* AKA NT_THREADS */
+#    define _Py_thread_local __declspec(thread)
+#  elif defined(__GNUC__)  /* includes clang */
+#    define _Py_thread_local __thread
+#  else
+#    error "no supported thread-local variable storage classifier"
 #  endif
 #endif
 
diff --git a/Lib/test/test_import/__init__.py b/Lib/test/test_import/__init__.py
index b71a36ec2f7aab4..95121debbbfa74e 100644
--- a/Lib/test/test_import/__init__.py
+++ b/Lib/test/test_import/__init__.py
@@ -3160,6 +3160,7 @@ def test_check_state_first(self):
     # Also, we test with a single-phase module that has global state,
     # which is shared by all interpreters.
 
+    @no_rerun(reason="module state is not cleared (see gh-140657)")
     @requires_subinterpreters
     def test_basic_multiple_interpreters_main_no_reset(self):
         # without resetting; already loaded in main interpreter
diff --git a/Python/import.c b/Python/import.c
index 45206b467938462..d4b574a8828dc5d 100644
--- a/Python/import.c
+++ b/Python/import.c
@@ -782,18 +782,13 @@ _PyImport_ClearModulesByIndex(PyInterpreterState *interp)
    substitute this (if the name actually matches).
 */
 
-#ifdef HAVE_THREAD_LOCAL
 _Py_thread_local const char *pkgcontext = NULL;
 # undef PKGCONTEXT
 # define PKGCONTEXT pkgcontext
-#endif
 
 const char *
 _PyImport_ResolveNameWithPackageContext(const char *name)
 {
-#ifndef HAVE_THREAD_LOCAL
-    PyMutex_Lock(&EXTENSIONS.mutex);
-#endif
     if (PKGCONTEXT != NULL) {
         const char *p = strrchr(PKGCONTEXT, '.');
         if (p != NULL && strcmp(name, p+1) == 0) {
@@ -801,23 +796,14 @@ _PyImport_ResolveNameWithPackageContext(const char *name)
             PKGCONTEXT = NULL;
         }
     }
-#ifndef HAVE_THREAD_LOCAL
-    PyMutex_Unlock(&EXTENSIONS.mutex);
-#endif
     return name;
 }
 
 const char *
 _PyImport_SwapPackageContext(const char *newcontext)
 {
-#ifndef HAVE_THREAD_LOCAL
-    PyMutex_Lock(&EXTENSIONS.mutex);
-#endif
     const char *oldcontext = PKGCONTEXT;
     PKGCONTEXT = newcontext;
-#ifndef HAVE_THREAD_LOCAL
-    PyMutex_Unlock(&EXTENSIONS.mutex);
-#endif
     return oldcontext;
 }
 
diff --git a/Python/pystate.c b/Python/pystate.c
index 2141e842a37d2f3..24681536797f94e 100644
--- a/Python/pystate.c
+++ b/Python/pystate.c
@@ -67,9 +67,6 @@ to avoid the expense of doing their own locking).
    For each of these functions, the GIL must be held by the current thread.
  */
 
-#ifndef HAVE_THREAD_LOCAL
-#  error "no supported thread-local variable storage classifier"
-#endif
 
 /* The attached thread state for the current thread. */
 _Py_thread_local PyThreadState *_Py_tss_tstate = NULL;
diff --git a/Python/traceback.c b/Python/traceback.c
index 2639d50aa667177..ef67368550aafb4 100644
--- a/Python/traceback.c
+++ b/Python/traceback.c
@@ -980,6 +980,52 @@ _Py_DumpASCII(int fd, PyObject *text)
     }
 }
 
+
+#ifdef MS_WINDOWS
+static void
+_Py_DumpWideString(int fd, wchar_t *str)
+{
+    Py_ssize_t size = wcslen(str);
+    int truncated;
+    if (MAX_STRING_LENGTH < size) {
+        size = MAX_STRING_LENGTH;
+        truncated = 1;
+    }
+    else {
+        truncated = 0;
+    }
+
+    for (Py_ssize_t i=0; i < size; i++) {
+        Py_UCS4 ch = str[i];
+        if (' ' <= ch && ch <= 126) {
+            /* printable ASCII character */
+            dump_char(fd, (char)ch);
+        }
+        else if (ch <= 0xff) {
+            PUTS(fd, "\\x");
+            _Py_DumpHexadecimal(fd, ch, 2);
+        }
+        else if (Py_UNICODE_IS_HIGH_SURROGATE(ch)
+                 && Py_UNICODE_IS_LOW_SURROGATE(str[i+1])) {
+            ch = Py_UNICODE_JOIN_SURROGATES(ch, str[i+1]);
+            i++;  // Skip the low surrogate character
+            PUTS(fd, "\\U");
+            _Py_DumpHexadecimal(fd, ch, 8);
+        }
+        else {
+            Py_BUILD_ASSERT(sizeof(wchar_t) == 2);
+            PUTS(fd, "\\u");
+            _Py_DumpHexadecimal(fd, ch, 4);
+        }
+    }
+
+    if (truncated) {
+        PUTS(fd, "...");
+    }
+}
+#endif
+
+
 /* Write a frame into the file fd: "File "xxx", line xxx in xxx".
 
    This function is signal safe. */
@@ -1149,20 +1195,15 @@ write_thread_name(int fd, PyThreadState *tstate)
         return;
     }
 
-    wchar_t *wname;
-    HRESULT hr = pGetThreadDescription(thread, &wname);
+    wchar_t *name;
+    HRESULT hr = pGetThreadDescription(thread, &name);
     if (!FAILED(hr)) {
-        char *name = _Py_EncodeLocaleRaw(wname, NULL);
-        if (name != NULL) {
-            size_t len = strlen(name);
-            if (len) {
-                PUTS(fd, " [");
-                (void)_Py_write_noraise(fd, name, len);
-                PUTS(fd, "]");
-            }
-            PyMem_RawFree(name);
+        if (name[0] != 0) {
+            PUTS(fd, " [");
+            _Py_DumpWideString(fd, name);
+            PUTS(fd, "]");
         }
-        LocalFree(wname);
+        LocalFree(name);
     }
     CloseHandle(thread);
 #endif