http client 实现 keep-alive 源码探究

前几天在分享"实现自己的wget"的时候，因为我们的请求是一次性的，http 头里设置的Connection: Close。在HTTP/1.1为了提升HTTP 1.0的网络性能，增加了keepalive的特性。浏览器在请求的时候都会加上Connection: Keep-Alive的头信息，是如何实现的呢？
我们知道在服务端（nginx）可以通过设置keepalive_timeout来控制连接保持时间，那么http连接的保持需要浏览器（客户端）支持吗？今天咱们一起来通过java.net.HttpURLConnection源码看看客户端是如何维护这些http连接的。

测试代码

package net.mengkang.demo;

import java.io.*;
import java.net.HttpURLConnection;
import java.net.URL;

public class Demo {
    public static void main(String[] args) throws IOException {
        test();
        test();
    }

    private static void test() throws IOException {
        URL url = new URL("http://static.mengkang.net/upload/image/2019/0921/1569075837628814.jpeg");

        HttpURLConnection connection = (HttpURLConnection) url.openConnection();
        connection.setRequestProperty("Charset", "UTF-8");
        connection.setRequestProperty("Connection", "Keep-Alive");
        connection.setRequestMethod("GET");
        connection.connect();

        BufferedInputStream bufferedInputStream = new BufferedInputStream(connection.getInputStream());

        File file = new File("./xxx.jpeg");
        OutputStream out = new FileOutputStream(file);
        int size;
        byte[] buf = new byte[1024];
        while ((size = bufferedInputStream.read(buf)) != -1) {
            out.write(buf, 0, size);
        }

        connection.disconnect();
    }
}

解析返回的头信息

当客户端从服务端获取返回的字节流时

connection.getInputStream()

HttpClient会对返回的头信息进行解析，我简化了摘取了最重要的逻辑代码

private boolean parseHTTPHeader(MessageHeader var1, ProgressSource var2, HttpURLConnection var3) throws IOException {
    String var15 = var1.findValue("Connection");
    ...
    if (var15 != null && var15.toLowerCase(Locale.US).equals("keep-alive")) {
        HeaderParser var11 = new HeaderParser(var1.findValue("Keep-Alive"));
        this.keepAliveConnections = var11.findInt("max", this.usingProxy ? 50 : 5);
        this.keepAliveTimeout = var11.findInt("timeout", this.usingProxy ? 60 : 5);
    }
    ...
}

是否需要保持长连接，是客户端申请，服务端决定，所以要以服务端返回的头信息为准。比如客户端发送的请求是Connection: Keep-Alive，服务端返回的是Connection: Close那也得以服务端为准。

客户端请求完成

当第一次执行时bufferedInputStream.read(buf)时，HttpClient会执行finished()方法

public void finished() {
    if (!this.reuse) {
        --this.keepAliveConnections;
        this.poster = null;
        if (this.keepAliveConnections > 0 && this.isKeepingAlive() && !this.serverOutput.checkError()) {
            this.putInKeepAliveCache();
        } else {
            this.closeServer();
        }

    }
}

加入到 http 长连接缓存

protected static KeepAliveCache kac = new KeepAliveCache();

protected synchronized void putInKeepAliveCache() {
    if (this.inCache) {
        assert false : "Duplicate put to keep alive cache";

    } else {
        this.inCache = true;
        kac.put(this.url, (Object)null, this);
    }
}

public class KeepAliveCache extends HashMap<KeepAliveKey, ClientVector> implements Runnable {
    ...
    public synchronized void put(URL var1, Object var2, HttpClient var3) {
        KeepAliveKey var5 = new KeepAliveKey(var1, var2); // var2 null
        ClientVector var6 = (ClientVector)super.get(var5);
        if (var6 == null) {
            int var7 = var3.getKeepAliveTimeout();
            var6 = new ClientVector(var7 > 0 ? var7 * 1000 : 5000);
            var6.put(var3);
            super.put(var5, var6);
        } else {
            var6.put(var3);
        }
    }
    ...
}

这里涉及了KeepAliveKey和ClientVector

class KeepAliveKey {
    private String protocol = null;
    private String host = null;
    private int port = 0;
    private Object obj = null;
}

设计这个对象呢，是因为只有protocol+host+port才能确定为同一个连接。所以用KeepAliveKey作为KeepAliveCache的key。
ClientVector则是一个栈，每次有同一个域下的请求都入栈。

class ClientVector extends Stack<KeepAliveEntry> {
    private static final long serialVersionUID = -8680532108106489459L;
    int nap;

    ClientVector(int var1) {
        this.nap = var1;
    }

    synchronized void put(HttpClient var1) {
        if (this.size() >= KeepAliveCache.getMaxConnections()) {
            var1.closeServer();
        } else {
            this.push(new KeepAliveEntry(var1, System.currentTimeMillis()));
        }
    }
    ...
}

“断开”连接

connection.disconnect();

如果是保持长连接的，实际只是关闭了一些流，socket 并没有关闭。

public void disconnect() {
...
      boolean var2 = var1.isKeepingAlive();
      if (var2) {
          var1.closeIdleConnection();
      }
...
}

public void closeIdleConnection() {
    HttpClient var1 = kac.get(this.url, (Object)null);
    if (var1 != null) {
        var1.closeServer();
    }
}

连接的复用

public static HttpClient New(URL var0, Proxy var1, int var2, boolean var3, HttpURLConnection var4) throws IOException {
    ...
    HttpClient var5 = null;
    if (var3) {
        var5 = kac.get(var0, (Object)null);
        ...
    }

    if (var5 == null) {
        var5 = new HttpClient(var0, var1, var2);
    } else {
        ...
        var5.url = var0;
    }

    return var5;
}

public class KeepAliveCache extends HashMap<KeepAliveKey, ClientVector> implements Runnable {
    ...
    public synchronized HttpClient get(URL var1, Object var2) {
        KeepAliveKey var3 = new KeepAliveKey(var1, var2);
        ClientVector var4 = (ClientVector)super.get(var3);
        return var4 == null ? null : var4.get();
    }
    ...
}

ClientVector取的时候则出栈，出栈过程中如果该连接已经超时，则关闭与服务端的连接，继续执行出栈操作。

class ClientVector extends Stack<KeepAliveEntry> {
    private static final long serialVersionUID = -8680532108106489459L;
    int nap;

    ClientVector(int var1) {
        this.nap = var1;
    }

    synchronized HttpClient get() {
        if (this.empty()) {
            return null;
        } else {
            HttpClient var1 = null;
            long var2 = System.currentTimeMillis();

            do {
                KeepAliveEntry var4 = (KeepAliveEntry)this.pop();
                if (var2 - var4.idleStartTime > (long)this.nap) {
                    var4.hc.closeServer();
                } else {
                    var1 = var4.hc;
                }
            } while(var1 == null && !this.empty());

            return var1;
        }
    }
    ...
}

这样就实现了客户端http连接的复用。

小结

存储结构如下

复用tcp的连接标准是protocol+host+port，客户端连接与服务端维持的连接数也不宜过多，HttpURLConnection默认只能存5个不同的连接，再多则直接断开连接（见上面HttpClient#finished方法），保持连接数过多对客户端和服务端都会增加不小的压力。
同时KeepAliveCache也每隔5秒钟扫描检测一次，清除过期的HttpClient。