scheduler: Improve power saving mode

* Remove possibility to temporarily disable small task packing.
 * When small task packing target gets overloaded continue packing
   threads on another core, but avoid migrating the already packed
   ones.

Scheduler still tends to needlessly migrate threads to another cores
when under heavier load, but it is now much better than before.

src/system/kernel/scheduler/power_saving.cpp

@@ -13,7 +13,6 @@
 using namespace Scheduler;
 
 
-static bigtime_t sDisableSmallTaskPacking;
 static int32 sSmallTaskCore;
 
 
@@ -32,54 +31,47 @@ has_cache_expired(Thread* thread)
 }
 
 
-static inline bool
-is_small_task_packing_enabled(void)
-{
-	if (sDisableSmallTaskPacking == -1)
-		return false;
-	return sDisableSmallTaskPacking < system_time();
-}
-
-
-static inline void
-disable_small_task_packing(void)
-{
-	ASSERT(!gSingleCore);
-
-	ASSERT(is_small_task_packing_enabled());
-	ASSERT(sSmallTaskCore == gCPUToCore[smp_get_current_cpu()]);
-
-	sDisableSmallTaskPacking = system_time() + kThreadQuantum * 100;
-	sSmallTaskCore = -1;
-}
-
-
-static inline bool
-is_task_small(Thread* thread)
-{
-	return thread->scheduler_data->load <= 200;
-}
-
-
 static void
 switch_to_mode(void)
 {
-	sDisableSmallTaskPacking = -1;
 	sSmallTaskCore = -1;
 }
 
 
+static bool
+try_small_task_packing(Thread* thread)
+{
+	int32 core = sSmallTaskCore;
+	return (core == -1 && gCoreLoadHeap->PeekMaximum() != NULL)
+		|| (core != -1
+			&& gCoreEntries[core].fLoad + thread->scheduler_data->load
+				< kHighLoad);
+}
+
+
+static int32
+choose_small_task_core(void)
+{
+	CoreEntry* candidate = gCoreLoadHeap->PeekMaximum();
+	if (candidate == NULL)
+		return sSmallTaskCore;
+
+	int32 core = candidate->fCoreID;
+	int32 smallTaskCore = atomic_test_and_set(&sSmallTaskCore, core, -1);
+	if (smallTaskCore == -1)
+		return core;
+	return smallTaskCore;
+}
+
+
 static int32
 choose_core(Thread* thread)
 {
 	CoreEntry* entry;
 
-	if (is_small_task_packing_enabled() && is_task_small(thread)
+	if (try_small_task_packing(thread)) {
-		&& gCoreLoadHeap->PeekMaximum() != NULL) {
 		// try to pack all threads on one core
-		if (sSmallTaskCore < 0)
+		entry = &gCoreEntries[choose_small_task_core()];
-			sSmallTaskCore = gCoreLoadHeap->PeekMaximum()->fCoreID;
-		entry = &gCoreEntries[sSmallTaskCore];
 	} else if (gCoreLoadHeap->PeekMinimum() != NULL) {
 		// run immediately on already woken core
 		entry = gCoreLoadHeap->PeekMinimum();
@@ -117,34 +109,28 @@ should_rebalance(Thread* thread)
 	int32 core = schedulerThreadData->previous_core;
 	CoreEntry* coreEntry = &gCoreEntries[core];
 
-	// If the thread produces more than 50% of the load, leave it here. In
-	// such situation it is better to move other threads away.
-	// Unless we are trying to pack small tasks here, in such case get rid
-	// of CPU hungry thread and continue packing.
-	if (schedulerThreadData->load >= coreEntry->fLoad / 2)
-		return is_small_task_packing_enabled() && sSmallTaskCore == core;
-
-	// All cores try to give us small tasks, check whether we have enough.
-	if (is_small_task_packing_enabled() && sSmallTaskCore == core) {
 	if (coreEntry->fLoad > kHighLoad) {
-			if (!is_task_small(thread))
+		if (sSmallTaskCore == core) {
+			SpinLocker coreLocker(gCoreHeapsLock);
+			CoreEntry* other = gCoreLoadHeap->PeekMaximum();
+				
+			if (other == NULL)
+				sSmallTaskCore = -1;
+			else if (coreEntry->fLoad - schedulerThreadData->load < kHighLoad)
 				return true;
-		} else if (coreEntry->fLoad > kVeryHighLoad)
+			else 
-			disable_small_task_packing();
+				sSmallTaskCore = other->fCoreID;
+			return coreEntry->fLoad > kVeryHighLoad;
+		}
+	} else if (coreEntry->fLoad < kHighLoad) {
+		int32 newCore = choose_small_task_core();
+		return newCore != core;
 	}
 
-	// Try small task packing.
+	CoreEntry* other = gCoreHighLoadHeap->PeekMinimum();
-	if (is_small_task_packing_enabled() && is_task_small(thread))
-		return sSmallTaskCore != core;
-
-	// No cpu bound threads - the situation is quite good. Make sure it
-	// won't get much worse...
-	SpinLocker coreLocker(gCoreHeapsLock);
-
-	CoreEntry* other = gCoreLoadHeap->PeekMinimum();
 	if (other == NULL)
-		other = gCoreHighLoadHeap->PeekMinimum();
+		return false;
-	return coreEntry->fLoad - other->fLoad >= kLoadDifference;
+	return coreEntry->fLoad - other->fLoad >= kLoadDifference / 2;
 }
 
 
@@ -175,12 +161,12 @@ pack_irqs(void)
 static void
 rebalance_irqs(bool idle)
 {
-	if (idle && !is_small_task_packing_enabled() && sSmallTaskCore != -1) {
+	if (idle && sSmallTaskCore != -1) {
 		pack_irqs();
 		return;
 	}
 
-	if (idle)
+	if (idle || sSmallTaskCore != -1)
 		return;
 
 	cpu_ent* cpu = get_cpu_struct();

src/system/kernel/scheduler/scheduler.cpp

@@ -1508,7 +1508,7 @@ _scheduler_init()
 			return result;
 	}
 
-#if 1
+#if 0
 	scheduler_set_operation_mode(SCHEDULER_MODE_LOW_LATENCY);
 #else
 	scheduler_set_operation_mode(SCHEDULER_MODE_POWER_SAVING);

src/system/kernel/scheduler/scheduler_common.h

@@ -40,11 +40,12 @@ const bigtime_t kMinimalWaitTime = kThreadQuantum / 4;
 
 const bigtime_t kCacheExpire = 100000;
 
+const int kLowLoad = kMaxLoad * 20 / 100;
 const int kTargetLoad = kMaxLoad * 55 / 100;
 const int kHighLoad = kMaxLoad * 70 / 100;
 const int kVeryHighLoad = (kMaxLoad + kHighLoad) / 2;
+
 const int kLoadDifference = kMaxLoad * 20 / 100;
-const int kLowLoad = kLoadDifference / 2;
 
 extern bool gSingleCore;