Fix get_stack_base in linux

In case of pthread_attr_getstack() returns np_stack_size = 1MB,
- calculated stack_base = stack_limit + stack_size(default: 4MB)
- real stack_base = stack_limit + np_stack_size(1MB)
stack_base is wrong because calculated with 4MB default stack size.

Due to this wrong stack_base, task_dispatcher::can_steal() can be always false.
It causes abnormal long loop in task_dispatcher::receive_or_steal_task().

Correct stack_size by pthread_attr_getstacksize() in linux.

Signed-off-by: Bongkyu Kim <[email protected]>

Author: bongkyu7-kim

src/tbb/governor.cpp

@@ -192,6 +192,22 @@ static std::uintptr_t get_stack_base(std::size_t stack_size) {
 #endif /* __TBB_USE_WINAPI */
 }
 
+static std::size_t get_stack_size(arena &a) {
+#if __linux__ && !__bg__
+    size_t np_stack_size = 0;
+    pthread_attr_t np_attr_stack;
+    if (0 == pthread_getattr_np(pthread_self(), &np_attr_stack)) {
+        if (0 == pthread_attr_getstacksize(&np_attr_stack, &np_stack_size)) {
+            __TBB_ASSERT( np_stack_size > 0, "stack size must be positive" );
+        }
+        pthread_attr_destroy(&np_attr_stack);
+    }
+    return np_stack_size > 0 ? np_stack_size : a.my_threading_control->worker_stack_size();
+#else
+    return a.my_threading_control->worker_stack_size();
+#endif
+}
+
 #if (_WIN32||_WIN64) && !__TBB_DYNAMIC_LOAD_ENABLED
 static void register_external_thread_destructor() {
     struct thread_destructor {
@@ -219,7 +235,7 @@ void governor::init_external_thread() {
     td.attach_arena(a, /*slot index*/ 0);
     __TBB_ASSERT(td.my_inbox.is_idle_state(false), nullptr);
 
-    stack_size = a.my_threading_control->worker_stack_size();
+    stack_size = get_stack_size(a);
     std::uintptr_t stack_base = get_stack_base(stack_size);
     task_dispatcher& task_disp = td.my_arena_slot->default_task_dispatcher();
     td.enter_task_dispatcher(task_disp, calculate_stealing_threshold(stack_base, stack_size));