异步
AsyncAppender
log4j2突出于其他日志的优势,异步日志实现。我们先从日志打印看进去。找到Logger,随便找一个log日志的方法。
public void debug(final Marker marker, final Message msg) {
logIfEnabled(FQCN, Level.DEBUG, marker, msg, msg != null ? msg.getThrowable() : null);
}
一路跟进
@PerformanceSensitive
// NOTE: This is a hot method. Current implementation compiles to 29 bytes of byte code.
// This is within the 35 byte MaxInlineSize threshold. Modify with care!
private void logMessageTrackRecursion(final String fqcn,
final Level level,
final Marker marker,
final Message msg,
final Throwable throwable) {
try {
incrementRecursionDepth(); // LOG4J2-1518, LOG4J2-2031
tryLogMessage(fqcn, level, marker, msg, throwable);
} finally {
decrementRecursionDepth();
}
}
可以看出这个在打日志之前做了调用次数的记录。跟进tryLogMessage,
@PerformanceSensitive
// NOTE: This is a hot method. Current implementation compiles to 26 bytes of byte code.
// This is within the 35 byte MaxInlineSize threshold. Modify with care!
private void tryLogMessage(final String fqcn,
final Level level,
final Marker marker,
final Message msg,
final Throwable throwable) {
try {
logMessage(fqcn, level, marker, msg, throwable);
} catch (final Exception e) {
// LOG4J2-1990 Log4j2 suppresses all exceptions that occur once application called the logger
handleLogMessageException(e, fqcn, msg);
}
}
继续跟进:
@Override
public void logMessage(final String fqcn, final Level level, final Marker marker, final Message message,
final Throwable t) {
final Message msg = message == null ? new SimpleMessage(Strings.EMPTY) : message;
final ReliabilityStrategy strategy = privateConfig.loggerConfig.getReliabilityStrategy();
strategy.log(this, getName(), fqcn, marker, level, msg, t);
}
这里可以看到在实际打日志的时候,会从config中获取打日志的策略,跟踪ReliabilityStrategy的创建,发现默认的实现类为DefaultReliabilityStrategy,跟进看实际打日志的方法
@Override
public void log(final Supplier<LoggerConfig> reconfigured, final String loggerName, final String fqcn, final Marker marker, final Level level,
final Message data, final Throwable t) {
loggerConfig.log(loggerName, fqcn, marker, level, data, t);
}
这里实际打日志的方法居然是交给一个config去实现的。。。感觉有点奇怪。。跟进看看
@PerformanceSensitive("allocation")
public void log(final String loggerName, final String fqcn, final Marker marker, final Level level,
final Message data, final Throwable t) {
List<Property> props = null;
if (!propertiesRequireLookup) {
props = properties;
} else {
if (properties != null) {
props = new ArrayList<>(properties.size());
final LogEvent event = Log4jLogEvent.newBuilder()
.setMessage(data)
.setMarker(marker)
.setLevel(level)
.setLoggerName(loggerName)
.setLoggerFqcn(fqcn)
.setThrown(t)
.build();
for (int i = 0; i < properties.size(); i++) {
final Property prop = properties.get(i);
final String value = prop.isValueNeedsLookup() // since LOG4J2-1575
? config.getStrSubstitutor().replace(event, prop.getValue()) //
: prop.getValue();
props.add(Property.createProperty(prop.getName(), value));
}
}
}
final LogEvent logEvent = logEventFactory.createEvent(loggerName, marker, fqcn, level, data, props, t);
try {
log(logEvent, LoggerConfigPredicate.ALL);
} finally {
// LOG4J2-1583 prevent scrambled logs when logging calls are nested (logging in toString())
ReusableLogEventFactory.release(logEvent);
}
}
可以清楚的看到try之前是在创建LogEvent,try里面做的才是真正的log(好tm累),一路跟进。
private void processLogEvent(final LogEvent event, LoggerConfigPredicate predicate) {
event.setIncludeLocation(isIncludeLocation());
if (predicate.allow(this)) {
callAppenders(event);
}
logParent(event, predicate);
}
接下来就是callAppender了,我们直接开始看AsyncAppender的append方法:
/**
* Actual writing occurs here.
*
* @param logEvent The LogEvent.
*/
@Override
public void append(final LogEvent logEvent) {
if (!isStarted()) {
throw new IllegalStateException("AsyncAppender " + getName() + " is not active");
}
final Log4jLogEvent memento = Log4jLogEvent.createMemento(logEvent, includeLocation);
InternalAsyncUtil.makeMessageImmutable(logEvent.getMessage());
if (!transfer(memento)) {
if (blocking) {
if (AbstractLogger.getRecursionDepth() > 1) { // LOG4J2-1518, LOG4J2-2031
// If queue is full AND we are in a recursive call, call appender directly to prevent deadlock
AsyncQueueFullMessageUtil.logWarningToStatusLogger();
logMessageInCurrentThread(logEvent);
} else {
// delegate to the event router (which may discard, enqueue and block, or log in current thread)
final EventRoute route = asyncQueueFullPolicy.getRoute(thread.getId(), memento.getLevel());
route.logMessage(this, memento);
}
} else {
error("Appender " + getName() + " is unable to write primary appenders. queue is full");
logToErrorAppenderIfNecessary(false, memento);
}
}
}
这里主要的步骤就是:
-
生成logEvent
-
将logEvent放入BlockingQueue,就是transfer方法
-
如果BlockingQueue满了则启用相应的策略
同样的,这里也有一个线程用来做异步消费的事情
private class AsyncThread extends Log4jThread {
private volatile boolean shutdown = false;
private final List<AppenderControl> appenders;
private final BlockingQueue<LogEvent> queue;
public AsyncThread(final List<AppenderControl> appenders, final BlockingQueue<LogEvent> queue) {
super("AsyncAppender-" + THREAD_SEQUENCE.getAndIncrement());
this.appenders = appenders;
this.queue = queue;
setDaemon(true);
}
@Override
public void run() {
while (!shutdown) {
LogEvent event;
try {
event = queue.take();
if (event == SHUTDOWN_LOG_EVENT) {
shutdown = true;
continue;
}
} catch (final InterruptedException ex) {
break; // LOG4J2-830
}
event.setEndOfBatch(queue.isEmpty());
final boolean success = callAppenders(event);
if (!success && errorAppender != null) {
try {
errorAppender.callAppender(event);
} catch (final Exception ex) {
// Silently accept the error.
}
}
}
// Process any remaining items in the queue.
LOGGER.trace("AsyncAppender.AsyncThread shutting down. Processing remaining {} queue events.",
queue.size());
int count = 0;
int ignored = 0;
while (!queue.isEmpty()) {
try {
final LogEvent event = queue.take();
if (event instanceof Log4jLogEvent) {
final Log4jLogEvent logEvent = (Log4jLogEvent) event;
logEvent.setEndOfBatch(queue.isEmpty());
callAppenders(logEvent);
count++;
} else {
ignored++;
LOGGER.trace("Ignoring event of class {}", event.getClass().getName());
}
} catch (final InterruptedException ex) {
// May have been interrupted to shut down.
// Here we ignore interrupts and try to process all remaining events.
}
}
LOGGER.trace("AsyncAppender.AsyncThread stopped. Queue has {} events remaining. "
+ "Processed {} and ignored {} events since shutdown started.", queue.size(), count, ignored);
}
/**
* Calls {@link AppenderControl#callAppender(LogEvent) callAppender} on all registered {@code AppenderControl}
* objects, and returns {@code true} if at least one appender call was successful, {@code false} otherwise. Any
* exceptions are silently ignored.
*
* @param event the event to forward to the registered appenders
* @return {@code true} if at least one appender call succeeded, {@code false} otherwise
*/
boolean callAppenders(final LogEvent event) {
boolean success = false;
for (final AppenderControl control : appenders) {
try {
control.callAppender(event);
success = true;
} catch (final Exception ex) {
// If no appender is successful the error appender will get it.
}
}
return success;
}
public void shutdown() {
shutdown = true;
if (queue.isEmpty()) {
queue.offer(SHUTDOWN_LOG_EVENT);
}
if (getState() == State.TIMED_WAITING || getState() == State.WAITING) {
this.interrupt(); // LOG4J2-1422: if underlying appender is stuck in wait/sleep/join/park call
}
}
}
直接看run方法:
-
阻塞获取logEvent
-
将logEvent分发出去
-
如果线程要退出了,将blockingQueue里面的event消费完在退出。
AsyncLogger
接从AsyncLogger的logMessage看进去:
public void logMessage(final String fqcn, final Level level, final Marker marker, final Message message,
final Throwable thrown) {
if (loggerDisruptor.isUseThreadLocals()) {
logWithThreadLocalTranslator(fqcn, level, marker, message, thrown);
} else {
// LOG4J2-1172: avoid storing non-JDK classes in ThreadLocals to avoid memory leaks in web apps
logWithVarargTranslator(fqcn, level, marker, message, thrown);
}
}
跟进logWithThreadLocalTranslator,
private void logWithThreadLocalTranslator(final String fqcn, final Level level, final Marker marker,
final Message message, final Throwable thrown) {
// Implementation note: this method is tuned for performance. MODIFY WITH CARE!
final RingBufferLogEventTranslator translator = getCachedTranslator();
initTranslator(translator, fqcn, level, marker, message, thrown);
initTranslatorThreadValues(translator);
publish(translator);
}
这里的逻辑很简单,就是将日志相关的信息转换成RingBufferLogEvent(RingBuffer是Disruptor的无所队列),然后将其发布到RingBuffer中。发布到RingBuffer中,那肯定也有消费逻辑。这时候有两种方式可以找到这个消费的逻辑。
-
找disruptor被使用的地方,然后查看,但是这样做会很容易迷惑
-
按照Log4j2的尿性,这种Logger都有对应的start方法,我们可以从start方法入手寻找
在start方法中,我们找到了一段代码:
final RingBufferLogEventHandler[] handlers = {new RingBufferLogEventHandler()};
disruptor.handleEventsWith(handlers);
直接看看这个RingBufferLogEventHandler的实现:
public class RingBufferLogEventHandler implements
SequenceReportingEventHandler<RingBufferLogEvent>, LifecycleAware {
private static final int NOTIFY_PROGRESS_THRESHOLD = 50;
private Sequence sequenceCallback;
private int counter;
private long threadId = -1;
@Override
public void setSequenceCallback(final Sequence sequenceCallback) {
this.sequenceCallback = sequenceCallback;
}
@Override
public void onEvent(final RingBufferLogEvent event, final long sequence,
final boolean endOfBatch) throws Exception {
event.execute(endOfBatch);
event.clear();
// notify the BatchEventProcessor that the sequence has progressed.
// Without this callback the sequence would not be progressed
// until the batch has completely finished.
if (++counter > NOTIFY_PROGRESS_THRESHOLD) {
sequenceCallback.set(sequence);
counter = 0;
}
}
/**
* Returns the thread ID of the background consumer thread, or {@code -1} if the background thread has not started
* yet.
* @return the thread ID of the background consumer thread, or {@code -1}
*/
public long getThreadId() {
return threadId;
}
@Override
public void onStart() {
threadId = Thread.currentThread().getId();
}
@Override
public void onShutdown() {
}
}
顺着接口找上去,发现一个接口:
/**
* Callback interface to be implemented for processing events as they become available in the {@link RingBuffer}
*
* @param <T> event implementation storing the data for sharing during exchange or parallel coordination of an event.
* @see BatchEventProcessor#setExceptionHandler(ExceptionHandler) if you want to handle exceptions propagated out of the handler.
*/
public interface EventHandler<T>
{
/**
* Called when a publisher has published an event to the {@link RingBuffer}
*
* @param event published to the {@link RingBuffer}
* @param sequence of the event being processed
* @param endOfBatch flag to indicate if this is the last event in a batch from the {@link RingBuffer}
* @throws Exception if the EventHandler would like the exception handled further up the chain.
*/
void onEvent(T event, long sequence, boolean endOfBatch) throws Exception;
}
通过注释可以发现,这个onEvent就是处理逻辑,回到RingBufferLogEventHandler的onEvent方法,发现里面有一个execute方法,跟进:
public void execute(final boolean endOfBatch) {
this.endOfBatch = endOfBatch;
asyncLogger.actualAsyncLog(this);
}
这个方法就是实际打日志了,AsyncLogger看起来还是比较简单的,只是使用了一个Disruptor。
插件化
之前在很多代码里面都可以看到
final PluginManager manager = new PluginManager(CATEGORY);
manager.collectPlugins(pluginPackages);
其实整个log4j2为了获得更好的扩展性,将自己的很多组件都做成了插件,然后在配置的时候去加载plugin。
跟进collectPlugins。
public void collectPlugins(final List<String> packages) {
final String categoryLowerCase = category.toLowerCase();
final Map<String, PluginType<?>> newPlugins = new LinkedHashMap<>();
// First, iterate the Log4j2Plugin.dat files found in the main CLASSPATH
Map<String, List<PluginType<?>>> builtInPlugins = PluginRegistry.getInstance().loadFromMainClassLoader();
if (builtInPlugins.isEmpty()) {
// If we didn't find any plugins above, someone must have messed with the log4j-core.jar.
// Search the standard package in the hopes we can find our core plugins.
builtInPlugins = PluginRegistry.getInstance().loadFromPackage(LOG4J_PACKAGES);
}
mergeByName(newPlugins, builtInPlugins.get(categoryLowerCase));
// Next, iterate any Log4j2Plugin.dat files from OSGi Bundles
for (final Map<String, List<PluginType<?>>> pluginsByCategory : PluginRegistry.getInstance().getPluginsByCategoryByBundleId().values()) {
mergeByName(newPlugins, pluginsByCategory.get(categoryLowerCase));
}
// Next iterate any packages passed to the static addPackage method.
for (final String pkg : PACKAGES) {
mergeByName(newPlugins, PluginRegistry.getInstance().loadFromPackage(pkg).get(categoryLowerCase));
}
// Finally iterate any packages provided in the configuration (note these can be changed at runtime).
if (packages != null) {
for (final String pkg : packages) {
mergeByName(newPlugins, PluginRegistry.getInstance().loadFromPackage(pkg).get(categoryLowerCase));
}
}
LOGGER.debug("PluginManager '{}' found {} plugins", category, newPlugins.size());
plugins = newPlugins;
}
处理逻辑如下:
-
从Log4j2Plugin.dat中加载所有的内置的plugin
-
然后将OSGi Bundles中的Log4j2Plugin.dat中的plugin加载进来
-
再加载传入的package路径中的plugin
-
最后加载配置中的plugin
逻辑还是比较简单的,但是我在看源码的时候发现了一个很有意思的东西,就是在加载log4j2 core插件的时候,也就是
PluginRegistry.getInstance().loadFromMainClassLoader()
这个方法,跟进到decodeCacheFiles:
private Map<String, List<PluginType<?>>> decodeCacheFiles(final ClassLoader loader) {
final long startTime = System.nanoTime();
final PluginCache cache = new PluginCache();
try {
final Enumeration<URL> resources = loader.getResources(PluginProcessor.PLUGIN_CACHE_FILE);
if (resources == null) {
LOGGER.info("Plugin preloads not available from class loader {}", loader);
} else {
cache.loadCacheFiles(resources);
}
} catch (final IOException ioe) {
LOGGER.warn("Unable to preload plugins", ioe);
}
final Map<String, List<PluginType<?>>> newPluginsByCategory = new HashMap<>();
int pluginCount = 0;
for (final Map.Entry<String, Map<String, PluginEntry>> outer : cache.getAllCategories().entrySet()) {
final String categoryLowerCase = outer.getKey();
final List<PluginType<?>> types = new ArrayList<>(outer.getValue().size());
newPluginsByCategory.put(categoryLowerCase, types);
for (final Map.Entry<String, PluginEntry> inner : outer.getValue().entrySet()) {
final PluginEntry entry = inner.getValue();
final String className = entry.getClassName();
try {
final Class<?> clazz = loader.loadClass(className);
final PluginType<?> type = new PluginType<>(entry, clazz, entry.getName());
types.add(type);
++pluginCount;
} catch (final ClassNotFoundException e) {
LOGGER.info("Plugin [{}] could not be loaded due to missing classes.", className, e);
} catch (final LinkageError e) {
LOGGER.info("Plugin [{}] could not be loaded due to linkage error.", className, e);
}
}
}
final long endTime = System.nanoTime();
final DecimalFormat numFormat = new DecimalFormat("#0.000000");
final double seconds = (endTime - startTime) * 1e-9;
LOGGER.debug("Took {} seconds to load {} plugins from {}",
numFormat.format(seconds), pluginCount, loader);
return newPluginsByCategory;
}
可以发现加载时候是从一个文件(PLUGIN_CACHE_FILE)获取所有要获取的plugin。看到这里的时候我有一个疑惑就是,为什么不用反射的方式直接去扫描,而是要从文件中加载进来,而且文件是写死的,很不容易扩展啊。然后我找了一下PLUGIN_CACHE_FILE这个静态变量的用处,发现了PluginProcessor这个类,这里用到了注解处理器。
/**
* Annotation processor for pre-scanning Log4j 2 plugins.
*/
@SupportedAnnotationTypes("org.apache.logging.log4j.core.config.plugins.*")
public class PluginProcessor extends AbstractProcessor {
// TODO: this could be made more abstract to allow for compile-time and run-time plugin processing
/**
* The location of the plugin cache data file. This file is written to by this processor, and read from by
* {@link org.apache.logging.log4j.core.config.plugins.util.PluginManager}.
*/
public static final String PLUGIN_CACHE_FILE =
"META-INF/org/apache/logging/log4j/core/config/plugins/Log4j2Plugins.dat";
private final PluginCache pluginCache = new PluginCache();
@Override
public boolean process(final Set<? extends TypeElement> annotations, final RoundEnvironment roundEnv) {
System.out.println("Processing annotations");
try {
final Set<? extends Element> elements = roundEnv.getElementsAnnotatedWith(Plugin.class);
if (elements.isEmpty()) {
System.out.println("No elements to process");
return false;
}
collectPlugins(elements);
writeCacheFile(elements.toArray(new Element[elements.size()]));
System.out.println("Annotations processed");
return true;
} catch (final IOException e) {
e.printStackTrace();
error(e.getMessage());
return false;
} catch (final Exception ex) {
ex.printStackTrace();
error(ex.getMessage());
return false;
}
}
}
(不太重要的方法省略)
我们可以看到在process方法中,PluginProcessor会先收集所有的Plugin,然后在写入文件。这样做的好处就是可以省去反射时候的开销。
然后我又看了一下Plugin这个注解,发现它的RetentionPolicy是RUNTIME,一般来说PluginProcessor是搭配RetentionPolicy.SOURCE,CLASS使用的,而且既然你把自己的Plugin扫描之后写在文件中了,RetentionPolicy就没有必要是RUNTIME了吧,这个是一个很奇怪的地方。
小结
总算是把Log4j2的代码看完了,发现它的设计理念很值得借鉴,为了灵活性,所有的东西都设计成插件式。互联网技术日益发展,各种中间件层出不穷,而作为工程师的我们更需要做的是去思考代码与代码之间的关系,毫无疑问的是,解耦是最具有美感的关系。