kernel: Begin abstracting out _sched_spinlock

In a uniprocessor system, _sched_spinlock may not need to be
held in all the same cases that it does in a multiprocessor
system. Removing those unnecessary usages can lead to better
performance on UP systems. In the case of uncontested taking
and giving of a semaphore, this can be as much as a +14%
performance gain.

Signed-off-by: Peter Mitsis <peter.mitsis@intel.com>

kernel/include/ksched.h

@@ -38,6 +38,12 @@ BUILD_ASSERT(K_LOWEST_APPLICATION_THREAD_PRIO
 #define Z_ASSERT_VALID_PRIO(prio, entry_point) __ASSERT((prio) == -1, "")
 #endif /* CONFIG_MULTITHREADING */
 
+#if (CONFIG_MP_MAX_NUM_CPUS == 1)
+#define LOCK_SCHED_SPINLOCK
+#else
+#define LOCK_SCHED_SPINLOCK   K_SPINLOCK(&_sched_spinlock)
+#endif
+
 extern struct k_spinlock _sched_spinlock;
 
 extern struct k_thread _thread_dummy;
@@ -159,11 +165,20 @@ static inline void unpend_thread_no_timeout(struct k_thread *thread)
 	thread->base.pended_on = NULL;
 }
 
+/*
+ * In a multiprocessor system, z_unpend_first_thread() must lock the scheduler
+ * spinlock _sched_spinlock. However, in a uniprocessor system, that is not
+ * necessary as the caller has already taken precautions (in the form of
+ * locking interrupts).
+ */
 static ALWAYS_INLINE struct k_thread *z_unpend_first_thread(_wait_q_t *wait_q)
 {
 	struct k_thread *thread = NULL;
 
-	K_SPINLOCK(&_sched_spinlock) {
+	__ASSERT_EVAL(, int key = arch_irq_lock(); arch_irq_unlock(key),
+		      !arch_irq_unlocked(key), "");
+
+	LOCK_SCHED_SPINLOCK {
 		thread = _priq_wait_best(&wait_q->waitq);
 		if (unlikely(thread != NULL)) {
 			unpend_thread_no_timeout(thread);

tests/benchmarks/sched/src/main.c

@@ -106,7 +106,10 @@ int main(void)
 	uint64_t tot = 0U;
 	uint32_t runs = 0U;
 
+	int key;
+
 	for (int i = 0; i < N_RUNS + N_SETTLE; i++) {
+		key = arch_irq_lock();
 		stamp(UNPENDING);
 		z_unpend_first_thread(&waitq);
 		stamp(UNPENDED_READYING);
@@ -121,6 +124,7 @@ int main(void)
 		 */
 		k_yield();
 		stamp(YIELDED);
+		arch_irq_unlock(key);
 
 		uint32_t avg, whole = stamps[4] - stamps[0];