kernel: Track load produced by interrupt handlers

headers/private/kernel/int.h

@@ -19,6 +19,16 @@
 struct kernel_args;
 
 
+enum interrupt_type {
+	INTERRUPT_TYPE_EXCEPTION,
+	INTERRUPT_TYPE_IRQ,
+	INTERRUPT_TYPE_LOCAL_IRQ,
+	INTERRUPT_TYPE_SYSCALL,
+	INTERRUPT_TYPE_ICI,
+	INTERRUPT_TYPE_UNKNOWN
+};
+
+
 #ifdef __cplusplus
 extern "C" {
 #endif
@@ -53,7 +63,8 @@ are_interrupts_enabled(void)
 #define restore_interrupts(status)	arch_int_restore_interrupts(status)
 
 
-status_t reserve_io_interrupt_vectors(long count, long startVector);
+status_t reserve_io_interrupt_vectors(long count, long startVector,
+	enum interrupt_type type);
 status_t allocate_io_interrupt_vectors(long count, long *startVector);
 void free_io_interrupt_vectors(long count, long startVector);
 

headers/private/kernel/load_tracking.h

@@ -0,0 +1,55 @@
+/*
+ * Copyright 2013 Paweł Dziepak, [email protected].
+ * Distributed under the terms of the MIT License.
+ */
+#ifndef _KERNEL_LOAD_TRACKING_H
+#define _KERNEL_LOAD_TRACKING_H
+
+
+const int32 kMaxLoad = 1000;
+const bigtime_t kLoadMeasureInterval = 50000;
+const bigtime_t kIntervalInaccuracy = kLoadMeasureInterval / 4;
+
+
+static int32
+compute_load(bigtime_t& measureTime, bigtime_t& measureActiveTime, int32& load)
+{
+	bigtime_t now = system_time();
+
+	if (measureTime == 0) {
+		measureTime = now;
+		return -1;
+	}
+
+	bigtime_t deltaTime = now - measureTime;
+
+	if (deltaTime < kLoadMeasureInterval)
+		return -1;
+
+	int32 oldLoad = load;
+	ASSERT(oldLoad >= 0 && oldLoad <= kMaxLoad);
+
+	int32 newLoad = measureActiveTime * kMaxLoad;
+	newLoad /= max_c(deltaTime, 1);
+	newLoad = max_c(min_c(newLoad, kMaxLoad), 0);
+
+	measureActiveTime = 0;
+	measureTime = now;
+
+	deltaTime += kIntervalInaccuracy;
+	int n = deltaTime / kLoadMeasureInterval;
+	ASSERT(n > 0);
+
+	if (n > 10)
+		load = newLoad;
+	else {
+		newLoad *= (1 << n) - 1;
+		load = (load + newLoad) / (1 << n);
+		ASSERT(load >= 0 && load <= kMaxLoad);
+	}
+
+	return oldLoad;
+}
+
+
+#endif	// _KERNEL_LOAD_TRACKING_H

src/system/kernel/arch/x86/32/descriptors.cpp

@@ -311,7 +311,7 @@ x86_descriptors_init(kernel_args* args)
 	set_trap_gate(0, 98, &trap98);	// for performance testing only
 	set_trap_gate(0, 99, &trap99);	// syscall interrupt
 
-	reserve_io_interrupt_vectors(2, 98);
+	reserve_io_interrupt_vectors(2, 98, INTERRUPT_TYPE_SYSCALL);
 
 	// configurable msi or msi-x interrupts
 	set_interrupt_gate(0, 100, &trap100);

src/system/kernel/arch/x86/arch_smp.cpp

@@ -90,7 +90,8 @@ arch_smp_init(kernel_args *args)
 
 	if (args->num_cpus > 1) {
 		// I/O interrupts start at ARCH_INTERRUPT_BASE, so all interrupts are shifted
-		reserve_io_interrupt_vectors(3, 0xfd - ARCH_INTERRUPT_BASE);
+		reserve_io_interrupt_vectors(3, 0xfd - ARCH_INTERRUPT_BASE,
+			INTERRUPT_TYPE_ICI);
 		install_io_interrupt_handler(0xfd - ARCH_INTERRUPT_BASE, &x86_ici_interrupt, NULL, B_NO_LOCK_VECTOR);
 		install_io_interrupt_handler(0xfe - ARCH_INTERRUPT_BASE, &x86_smp_error_interrupt, NULL, B_NO_LOCK_VECTOR);
 		install_io_interrupt_handler(0xff - ARCH_INTERRUPT_BASE, &x86_spurious_interrupt, NULL, B_NO_LOCK_VECTOR);

src/system/kernel/arch/x86/ioapic.cpp

@@ -774,7 +774,7 @@ ioapic_init(kernel_args* args)
 	current = sIOAPICs;
 	while (current != NULL) {
 		reserve_io_interrupt_vectors(current->max_redirection_entry + 1,
-			current->global_interrupt_base);
+			current->global_interrupt_base, INTERRUPT_TYPE_IRQ);
 		current = current->next;
 	}
 

src/system/kernel/arch/x86/pic.cpp

@@ -231,7 +231,7 @@ pic_init()
 
 	TRACE(("PIC level trigger mode: 0x%08lx\n", sLevelTriggeredInterrupts));
 
-	reserve_io_interrupt_vectors(16, 0);
+	reserve_io_interrupt_vectors(16, 0, INTERRUPT_TYPE_EXCEPTION);
 
 	// make the pic controller the current one
 	arch_int_set_interrupt_controller(picController);

src/system/kernel/arch/x86/timers/x86_apic.cpp

@@ -106,7 +106,8 @@ apic_timer_init(struct kernel_args *args)
 
 	sApicTicsPerSec = args->arch_args.apic_time_cv_factor;
 
-	reserve_io_interrupt_vectors(1, 0xfb - ARCH_INTERRUPT_BASE);
+	reserve_io_interrupt_vectors(1, 0xfb - ARCH_INTERRUPT_BASE,
+		INTERRUPT_TYPE_LOCAL_IRQ);
 	install_io_interrupt_handler(0xfb - ARCH_INTERRUPT_BASE,
 		&apic_timer_interrupt, NULL, B_NO_LOCK_VECTOR);
 

src/system/kernel/int.cpp

@@ -20,6 +20,7 @@
 #include <arch/int.h>
 #include <boot/kernel_args.h>
 #include <elf.h>
+#include <load_tracking.h>
 #include <util/AutoLock.h>
 #include <util/kqueue.h>
 #include <smp.h>
@@ -51,6 +52,13 @@ struct io_vector {
 	spinlock			vector_lock;
 	int32				enable_count;
 	bool				no_lock_vector;
+	interrupt_type		type;
+
+	spinlock			load_lock;
+	bigtime_t			last_measure_time;
+	bigtime_t			last_measure_active;
+	int32				load;
+
 #if DEBUG_INTERRUPTS
 	int64				handled_count;
 	int64				unhandled_count;
@@ -116,6 +124,29 @@ dump_int_statistics(int argc, char **argv)
 #endif
 
 
+static int
+dump_int_load(int argc, char** argv)
+{
+	static const char* typeNames[]
+		= { "exception", "irq", "local irq", "syscall", "ici", "unknown" };
+
+	for (int i = 0; i < NUM_IO_VECTORS; i++) {
+		if (!B_SPINLOCK_IS_LOCKED(&sVectors[i].vector_lock)
+			&& sVectors[i].handler_list == NULL
+			&& sVectors[i].enable_count == 0)
+			continue;
+
+		kprintf("int %3d, type %s, enabled %" B_PRId32 ", load %" B_PRId32
+			"%%%s\n", i, typeNames[min_c(sVectors[i].type,
+					INTERRUPT_TYPE_UNKNOWN)],
+			sVectors[i].enable_count, sVectors[i].load / 10,
+			B_SPINLOCK_IS_LOCKED(&sVectors[i].vector_lock) ? ", ACTIVE" : "");
+	}
+
+	return 0;
+}
+
+
 //	#pragma mark - private kernel API
 
 
@@ -145,6 +176,13 @@ int_init_post_vm(kernel_args* args)
 		B_INITIALIZE_SPINLOCK(&sVectors[i].vector_lock);
 		sVectors[i].enable_count = 0;
 		sVectors[i].no_lock_vector = false;
+		sVectors[i].type = INTERRUPT_TYPE_UNKNOWN;
+
+		B_INITIALIZE_SPINLOCK(&sVectors[i].load_lock);
+		sVectors[i].last_measure_time = 0;
+		sVectors[i].last_measure_active = 0;
+		sVectors[i].load = 0;
+
 #if DEBUG_INTERRUPTS
 		sVectors[i].handled_count = 0;
 		sVectors[i].unhandled_count = 0;
@@ -159,6 +197,9 @@ int_init_post_vm(kernel_args* args)
 		"list interrupt statistics");
 #endif
 
+	add_debugger_command("int_load", &dump_int_load,
+		"list interrupt usage statistics");
+
 	return arch_int_init_post_vm(args);
 }
 
@@ -179,6 +220,17 @@ int_init_post_device_manager(kernel_args* args)
 }
 
 
+static void
+update_int_load(int i)
+{
+	if (!try_acquire_spinlock(&sVectors[i].load_lock))
+		return;
+	compute_load(sVectors[i].last_measure_time, sVectors[i].last_measure_active,
+		sVectors[i].load);
+	release_spinlock(&sVectors[i].load_lock);
+}
+
+
 /*!	Actually process an interrupt via the handlers registered for that
 	vector (IRQ).
 */
@@ -189,6 +241,8 @@ int_io_interrupt_handler(int vector, bool levelTriggered)
 	struct io_handler* io;
 	bool handled = false;
 
+	bigtime_t start = system_time();
+
 	if (!sVectors[vector].no_lock_vector)
 		acquire_spinlock(&sVectors[vector].vector_lock);
 
@@ -267,6 +321,12 @@ int_io_interrupt_handler(int vector, bool levelTriggered)
 	if (!sVectors[vector].no_lock_vector)
 		release_spinlock(&sVectors[vector].vector_lock);
 
+	SpinLocker locker(sVectors[vector].load_lock);
+	sVectors[vector].last_measure_active += system_time() - start;
+	locker.Unlock();
+
+	update_int_load(vector);
+
 	if (levelTriggered)
 		return status;
 
@@ -435,7 +495,7 @@ remove_io_interrupt_handler(long vector, interrupt_handler handler, void *data)
 	vectors using allocate_io_interrupt_vectors() instead.
 */
 status_t
-reserve_io_interrupt_vectors(long count, long startVector)
+reserve_io_interrupt_vectors(long count, long startVector, interrupt_type type)
 {
 	MutexLocker locker(&sIOInterruptVectorAllocationLock);
 
@@ -448,6 +508,7 @@ reserve_io_interrupt_vectors(long count, long startVector)
 			return B_BUSY;
 		}
 
+		sVectors[startVector + i].type = type;
 		sAllocatedIOInterruptVectors[startVector + i] = true;
 	}
 

src/system/kernel/scheduler/scheduler.cpp

@@ -23,6 +23,7 @@
 #include <kernel.h>
 #include <kscheduler.h>
 #include <listeners.h>
+#include <load_tracking.h>
 #include <scheduler_defs.h>
 #include <smp.h>
 #include <thread.h>
@@ -61,10 +62,9 @@ const bigtime_t kMinimalWaitTime = kThreadQuantum / 4;
 
 const bigtime_t kCacheExpire = 100000;
 
-const int kTargetLoad = 550;
+const int kTargetLoad = kMaxLoad * 55 / 100;
-const int kHighLoad = 700;
+const int kHighLoad = kMaxLoad * 70 / 100;
-const int kMaxLoad = 1000;
+const int kLoadDifference = kMaxLoad * 20 / 100;
-const int kLoadDifference = 200;
 
 static bigtime_t sDisableSmallTaskPacking;
 static int32 sSmallTaskCore;
@@ -924,50 +924,6 @@ should_rebalance(Thread* thread)
 }
 
 
-static inline int
-compute_load(bigtime_t& measureTime, bigtime_t& measureActiveTime, int32& load)
-{
-	const bigtime_t kLoadMeasureInterval = 50000;
-	const bigtime_t kIntervalInaccuracy = kLoadMeasureInterval / 4;
-
-	bigtime_t now = system_time();
-
-	if (measureTime == 0) {
-		measureTime = now;
-		return -1;
-	}
-
-	bigtime_t deltaTime = now - measureTime;
-
-	if (deltaTime < kLoadMeasureInterval)
-		return -1;
-
-	int oldLoad = load;
-	ASSERT(oldLoad >= 0 && oldLoad <= kMaxLoad);
-
-	int newLoad = measureActiveTime * kMaxLoad;
-	newLoad /= max_c(deltaTime, 1);
-	newLoad = max_c(min_c(newLoad, kMaxLoad), 0);
-
-	measureActiveTime = 0;
-	measureTime = now;
-
-	deltaTime += kIntervalInaccuracy;
-	int n = deltaTime / kLoadMeasureInterval;
-	ASSERT(n > 0);
-
-	if (n > 10)
-		load = newLoad;
-	else {
-		newLoad *= (1 << n) - 1;
-		load = (load + newLoad) / (1 << n);
-		ASSERT(load >= 0 && load <= kMaxLoad);
-	}
-
-	return oldLoad;
-}
-
-
 static inline void
 compute_cpu_load(int32 cpu)
 {