java - 毕竟 log4j 不是线程安全的吗？

Question

(这是在 SLES11、Java 7、Tomcat 6、log4j-1.2.16 上)

我们使用 log4j 将不同的内容写入不同的日志文件。我继承了这段代码，所以无论好坏，通用结构都暂时保留在这里。

记录器将创建两个日志文件:main.log 和 stats.log。特定的统计消息通过单独的调用记录到两个记录器(您将在下面看到)，并且一大堆其他内容记录到主日志。

因此，在我们的所有代码中，您会看到类似 Log.logMain(someMessageToLog); 的内容。在我们的代码(由多个线程执行)的一个地方有以下内容:

String statsMessage = createStatsMessage();
Log.logMain(statsMessage);
Log.logStats(statsMessage);

主记录器的名称是main，统计记录器的名称是stats。问题是，有时在重负载下，我们会在 main.log 中看到包含字符串 stats INFO 的行。 main.log 中的所有内容都应该只有 main INFO，因为这是唯一记录到该文件的记录器，而且我们在某些行中看到了混合输出。这似乎是一个线程安全问题，但 log4j 文档说 log4j 是线程安全的。这是我的意思的一个例子:

2012-03-21 16:01:34,7742012-03-21 16:01:34,774| | stats main INFO   [INFO  http-8080-18]:  [message redacted]. 
2012-03-21 16:01:36,380| main 2012-03-21 16:01:36,380| INFO   [stats INFO  http-8080-15]: [message redacted]. 
2012-03-21 16:01:37,465| main INFO  2012-03-21 16:01:37,465 [| stats http-8080-1]: [message redacted]. 

这是 Log 类(精简后只显示有问题的记录器——实际上还有很多其他记录器，所有设置都与这些类似):

import org.apache.log4j.*;

import java.io.IOException;

final public class Log
{
    private static final String LOG_IDENTIFIER_MAINLOG = "main";
    private static final String LOG_IDENTIFIER_STATSLOG = "stats";

    private static final String MAIN_FILENAME = "/var/log/app_main.log";
    private static final String STATS_FILENAME = "/var/log/app_stats.log";

    private static final int BACKUP_INDEX = 40;
    private static final String BACKUP_SIZE = "10MB";

    private static final PatternLayout COMMON_LAYOUT =
        new PatternLayout("%d| %c %-6p [%t]: %m.%n");

    private static Logger mainLogger;
    private static Logger statsLogger;

    public static void init() {
        init(MAIN_FILENAME, STATS_FILENAME);
    }

    public static void init(String mainLogFilename,
                            String statsLogFilename) {
        mainLogger = initializeMainLogger(mainLogFilename);
        statsLogger = initializeStatsLogger(statsLogFilename);
    }

    public static void logMain(String message) {
        if (mainLogger != null) {
            mainLogger.info(message);
        }
    }

    public static void logStats(String message) {
        if (statsLogger != null) {
            statsLogger.info(message);
        }
    }

    private static Logger getLogger(String loggerIdentifier) {
        Logger logger = Logger.getLogger(loggerIdentifier);
        logger.setAdditivity(false);
        return logger;
    }

    private static boolean addFileAppender(Logger logger,
                                           String logFilename,
                                           int    maxBackupIndex,
                                           String maxSize) {
        try {
            RollingFileAppender appender =
                new RollingFileAppender(COMMON_LAYOUT, logFilename);
            appender.setMaxBackupIndex(maxBackupIndex);
            appender.setMaxFileSize(maxSize);
            logger.addAppender(appender);
        }
        catch (IOException ex) {
            ex.printStackTrace();
            return false;
        }
        return true;
    }

    private static Logger initializeMainLogger(String filename) {
        Logger logger = getLogger(LOG_IDENTIFIER_MAINLOG);
        addFileAppender(logger, filename, BACKUP_INDEX, BACKUP_SIZE);
        logger.setLevel(Level.INFO);
        return logger;
    }

    private static Logger initializeStatsLogger(String filename) {
        Logger logger = getLogger(LOG_IDENTIFIER_STATSLOG);
        addFileAppender(logger, filename, BACKUP_INDEX, BACKUP_SIZE);
        logger.setLevel(Level.INFO);
        return logger;
    }

}

---

更新:

这里有一个小程序(至少对我而言)可以重现上述 Log 类的问题:

final public class Stress
{
    public static void main(String[] args) throws Exception {
        if (args.length != 2) {
            Log.init();
        }
        else {
            Log.init(args[0], args[1]);
        }

        for (;;) {
            // I know Executors are preferred, but this
            // is a quick & dirty test program
            Thread t = new Thread(new TestLogging());
            t.start();
        }
    }

    private static final class TestLogging implements Runnable
    {
        private static int counter = 0;

        @Override
        public void run() {
            String msg = new StringBuilder("Count is: ")
                .append(counter++).toString();

            Log.logMain(msg);
            Log.logStats(msg);

            try {
                Thread.sleep(1);
            }
            catch (InterruptedException e) {
                Log.logMain(e.getMessage());
            }
        }
    }
}

以及日志中的一些示例输出:

$ grep stats main.log    
2012-03-23 15:30:35,919| stats 2012-03-23 15:30:35,919| main INFO  INFO   [ [Thread-313037]: Thread-313036]: Count is: 312987.
2012-03-23 15:30:35,929| stats INFO   [Thread-313100]: Count is: 313050.
2012-03-23 15:30:35,937| stats INFO   [Thread-313168]: Count is: 313112.
2012-03-23 15:30:35,945| stats INFO   [Thread-313240]: Count is: 313190.
2012-03-23 15:30:35,946| stats INFO   [Thread-313251]: Count is: 313201.
2012-03-23 15:30:35,949| stats INFO   [2012-03-23 15:30:35,949| main INFO  Thread-313281]: Count is: 313231.
2012-03-23 15:30:35,954| stats INFO   [Thread-313331]: Count is: 313281.
2012-03-23 15:30:35,956| 2012-03-23 15:30:35,956stats | main INFOINFO   [   [Thread-313356]: Count is: 313306.
2012-03-23 15:30:35,9562012-03-23 15:30:35,956| main | INFO  stats  [INFOThread-313359]:   Count is: 313309.
2012-03-23 15:30:35,962| stats INFO  2012-03-23 15:30:35,962| main INFO   [Thread-313388]:  [Count is: 313338.

和

$ grep main stats.log
2012-03-23 15:30:35,913| 2012-03-23 15:30:35,913| main INFO   [Thread-312998]: Count is: 312948.
2012-03-23 15:30:35,915| main INFO   [Thread-313014]: Count is: 312964.
2012-03-23 15:30:35,919| stats 2012-03-23 15:30:35,919| main INFO  INFO   [ [Thread-313037]: Thread-313036]: Count is: 312987.
2012-03-23 15:30:35,931| main INFO   [Thread-313116]: Count is: 313066.
2012-03-23 15:30:35,947| main INFO   [2012-03-23 15:30:35,947Thread-313264]: | Count is: 313214.
2012-03-23 15:30:35,949| stats INFO   [2012-03-23 15:30:35,949| main INFO  Thread-313281]: Count is: 313231.
2012-03-23 15:30:35,956| 2012-03-23 15:30:35,956stats | main INFOINFO   [   [Thread-313356]: Count is: 313306.
2012-03-23 15:30:35,9562012-03-23 15:30:35,956| main | INFO  stats  [INFOThread-313359]:   Count is: 313309.
2012-03-23 15:30:35,962| stats INFO  2012-03-23 15:30:35,962| main INFO   [Thread-313388]:  [Count is: 313338.

就其值(value)而言，在一个 145516 行的 main.log 文件中，“stats”出现了 2452 次。所以这并不少见，但也不是一直都在发生(当然，这个测试非常极端)。

Answer 1

http://logging.apache.org/log4j/1.2/apidocs/org/apache/log4j/PatternLayout.html

你在两个 appender 之间共享 PatternLayout，根据上面的 API 链接:

已知此代码存在同步和其他问题，这些问题在 org.apache.log4j.EnhancedPatternLayout 中不存在。应优先使用 EnhancedPatternLayout 而不是 PatternLayout。 EnhancedPatternLayout 分布在 log4j 附加组件中。

因此为每个appender创建一个新的PatternLayout