Convert important polling loops to spin at lowest SMT priority

The pattern of calling cpu_relax() inside a polling loop does
not suit the powerpc SMT priority instructions. Prefrred is to
set a low priority then spin until break condition is reached,
then restore priority.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
[stewart@linux.vnet.ibm.com: fixup lpc-uart wait_tx_room() and unit test]
Signed-off-by: Stewart Smith <stewart@linux.vnet.ibm.com>

core/hmi.c

@@ -599,6 +599,7 @@ static void wait_for_subcore_threads(void)
 {
 	uint64_t timeout = 0;
 
+	smt_lowest();
 	while (!(*(this_cpu()->core_hmi_state_ptr) & HMI_STATE_CLEANUP_DONE)) {
 		/*
 		 * We use a fixed number of TIMEOUT_LOOPS rather
@@ -616,8 +617,9 @@ static void wait_for_subcore_threads(void)
 			prlog(PR_DEBUG, "HMI: TB pre-recovery timeout\n");
 			break;
 		}
-		cpu_relax();
+		barrier();
 	}
+	smt_medium();
 }
 
 /*

core/lock.c

@@ -93,7 +93,10 @@ void lock(struct lock *l)
 	for (;;) {
 		if (try_lock(l))
 			break;
-		cpu_relax();
+		smt_lowest();
+		while (l->lock_val)
+			barrier();
+		smt_medium();
 	}
 }
 

core/test/run-timer.c

@@ -7,6 +7,8 @@
 
 #define mftb()	(stamp)
 #define sync()
+#define smt_lowest()
+#define smt_medium()
 
 static uint64_t stamp, last;
 struct lock;

core/timer.c

@@ -45,7 +45,10 @@ static void __sync_timer(struct timer *t)
 
 	while (t->running) {
 		unlock(&timer_lock);
-		cpu_relax();
+		smt_lowest();
+		while (t->running)
+			barrier();
+		smt_medium();
 		/* Should we call the pollers here ? */
 		lock(&timer_lock);
 	}

hw/lpc-uart.c

@@ -115,10 +115,16 @@ static void uart_check_tx_room(void)
 
 static void uart_wait_tx_room(void)
 {
-	while(!tx_room) {
+	while (!tx_room) {
 		uart_check_tx_room();
-		if (!tx_room)
-			cpu_relax();
+		if (!tx_room) {
+			smt_lowest();
+			do {
+				barrier();
+				uart_check_tx_room();
+			} while (!tx_room);
+			smt_medium();
+		}
 	}
 }