Index: core/stackless_tstate.h
===================================================================
--- core/stackless_tstate.h	(revision 60407)
+++ core/stackless_tstate.h	(working copy)
@@ -36,6 +36,7 @@
 #endif
 	/* number of nested interpreters (1.0/2.0 merge) */
 	int nesting_level;
+	int globaltrace; /* tracing is global for thread */
 } PyStacklessState;
 
 
@@ -53,7 +54,8 @@
 	tstate->st.main = NULL; \
 	tstate->st.current = NULL; \
 	tstate->st.runcount = 0; \
-	tstate->st.nesting_level = 0;
+	tstate->st.nesting_level = 0; \
+	tstate->st.globaltrace = 0;
 
 /* note that the scheduler knows how to zap. It checks if it is in charge
    for this tstate and then clears everything. This will not work if
Index: core/stacklesseval.c
===================================================================
--- core/stacklesseval.c	(revision 60407)
+++ core/stacklesseval.c	(working copy)
@@ -367,8 +367,14 @@
 {
 	PyThreadState *ts = PyThreadState_Get();
 
-	if (ts->st.main == NULL)
-		initialize_main_and_current();
+	if (ts->st.main == NULL) {
+		if (initialize_main_and_current()) {
+			PyObject *s = PyString_FromString("tasklet cleanup");
+			PyErr_WriteUnraisable(s);
+			Py_XDECREF(s);
+			return;
+		}
+	}
 	slp_kill_tasks_with_stacks(allthreads ? NULL : ts);
 }
 
Index: module/scheduling.c
===================================================================
--- module/scheduling.c	(revision 60407)
+++ module/scheduling.c	(working copy)
@@ -321,25 +321,30 @@
 	int ret;
 
 	ts->exc_type = ts->exc_value = ts->exc_traceback = NULL;
-	ts->c_profilefunc = ts->c_tracefunc = NULL;
-	ts->c_profileobj = ts->c_traceobj = NULL;
-	ts->use_tracing = ts->tracing = 0;
 
-	/* note that trace/profile are set without ref */
-	Py_XINCREF(c_profileobj);
-	Py_XINCREF(c_traceobj);
+	if (!ts->st.globaltrace) {
+		ts->c_profilefunc = ts->c_tracefunc = NULL;
+		ts->c_profileobj = ts->c_traceobj = NULL;
+		ts->use_tracing = ts->tracing = 0;
 
+		/* note that trace/profile are set without ref */
+		Py_XINCREF(c_profileobj);
+		Py_XINCREF(c_traceobj);
+	}
+
 	ret = slp_transfer(cstprev, cst, prev);
 
-	ts->tracing = tracing;
-	ts->use_tracing = use_tracing;
+	if (!ts->st.globaltrace) {
+		ts->tracing = tracing;
+		ts->use_tracing = use_tracing;
 
-	ts->c_profilefunc = c_profilefunc;
-	ts->c_tracefunc = c_tracefunc;
-	ts->c_profileobj = c_profileobj;
-	ts->c_traceobj = c_traceobj;
-	Py_XDECREF(c_profileobj);
-	Py_XDECREF(c_traceobj);
+		ts->c_profilefunc = c_profilefunc;
+		ts->c_tracefunc = c_tracefunc;
+		ts->c_profileobj = c_profileobj;
+		ts->c_traceobj = c_traceobj;
+		Py_XDECREF(c_profileobj);
+		Py_XDECREF(c_traceobj);
+	}
 
 	ts->exc_type = exc_type;
 	ts->exc_value = exc_value;
@@ -850,7 +855,7 @@
 		ts->exc_type = ts->exc_value =
 			       ts->exc_traceback = NULL;
 	}
-	if (ts->use_tracing || ts->tracing) {
+	if (!ts->st.globaltrace && (ts->use_tracing || ts->tracing)) {
 		/* build a shadow frame */
 		PyCFrameObject *f = slp_cframe_new(restore_tracing, 1);
 		if (f == NULL)
Index: module/stacklessmodule.c
===================================================================
--- module/stacklessmodule.c	(revision 60407)
+++ module/stacklessmodule.c	(working copy)
@@ -853,6 +853,26 @@
 }
 
 static PyObject *
+slpmodule_get_globaltrace(PySlpModuleObject *mod, void *context)
+{
+	PyThreadState *ts = PyThreadState_GET();
+	PyObject *result = ts->st.globaltrace ? Py_True : Py_False;
+	Py_INCREF(result);
+	return result;
+}
+
+static int
+slpmodule_set_globaltrace(PySlpModuleObject *mod, PyObject *val, void *context)
+{
+	PyThreadState *ts = PyThreadState_GET();
+	int r = PyObject_IsTrue(val);
+	if (r == -1)
+		return r;
+	ts->st.globaltrace = r;
+	return 0;
+}
+
+static PyObject *
 slpmodule_getdebug(PySlpModuleObject *mod)
 {
 #ifdef _DEBUG
@@ -932,6 +952,9 @@
 	 uncollectables__doc__},
 	{"threads", (getter)slpmodule_getthreads, NULL,
 	 "a list of all thread ids, starting with main."},
+	{"globaltrace", (getter)slpmodule_get_globaltrace,
+		(setter)slpmodule_set_globaltrace,
+	"is tracing or profiling global for the thread"},
 	{0},
 };