On Wed, Jun 02, 2021 at 10:06:58AM -0400, Mathieu Desnoyers wrote:

----- On Jun 2, 2021, at 9:12 AM, Peter Zijlstra peterz@infradead.org wrote:

quoted

@@ -134,14 +134,14 @@ struct task_group;

	do {							\
		WARN_ON_ONCE(is_special_task_state(state_value));\
		current->task_state_change = _THIS_IP_;		\
-		current->state = (state_value);			\
+		WRITE_ONCE(current->__state, (state_value));	\
	} while (0)

Why not introduce set_task_state(p) and get_task_state(p) rather than sprinkle
READ_ONCE() and WRITE_ONCE() all over the kernel ?

set_task_state() is fundamentally unsound, there's very few sites that
_set_ state on anything other than current, and those sites are super
tricky, eg. ptrace.

Having get_task_state() would seem to suggest it's actually a sane thing
to do, it's not really. Inspecting remote state is full of races, and
some of that really wants cleaning up, but that's for another day.