cobalt/sched: dovetail: prevent early scheduling on uninit runqueues

Make sure not to raise XNSCHED when setting up the root thread, so
that we can't start rescheduling from irq_exit_pipeline() before all
CPUs have their runqueue fully built.

At this chance, add the missing tracepoint reporting a remote
scheduling request.
Signed-off-by: Philippe Gerum <rpm@xenomai.org>
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

kernel/cobalt/dovetail/sched.c

@@ -91,6 +91,7 @@ void pipeline_clear_mayday(void) /* May solely affect current. */
 
 irqreturn_t pipeline_reschedule_ipi_handler(int irq, void *dev_id)
 {
+	trace_cobalt_schedule_remote(xnsched_current());
 
 	/* Will reschedule from irq_exit_pipeline(). */
 

kernel/cobalt/sched.c

@@ -465,7 +465,15 @@ int xnsched_set_policy(struct xnthread *thread,
 	if (xnthread_test_state(thread, XNREADY))
 		xnsched_enqueue(thread);
 
-	if (!xnthread_test_state(thread, XNDORMANT))
+	/*
+	 * Make sure not to raise XNSCHED when setting up the root
+	 * thread, so that we can't start rescheduling on interrupt
+	 * exit before all CPUs have their runqueue fully
+	 * built. Filtering on XNROOT here is correct because the root
+	 * thread enters the idle class once as part of the runqueue
+	 * setup process and never leaves it afterwards.
+	 */
+	if (!xnthread_test_state(thread, XNDORMANT|XNROOT))
 		xnsched_set_resched(thread->sched);
 
 	return 0;