wolfcstech | 作者
承香墨影 | 校對
https://www.wolfcstech.com/2017/02/23/OkHttp實現分析之Websocket | 原文
前段時間寫了篇文章,講解了如何使用 OkHttp 實現 WebSocket 通信,但是隻停留在一些使用和理論上,分享出來各項數據還不錯。本來想繼續寫一篇 OkHttp 關於 WebSocket 源碼的分析文章。但搜索了一下,發現有一個個人博客寫的文章非常詳細,我就不再重複輸出了,今天就給你帶來這篇優秀的文章,希望對大家有幫助。
如果只對利用 OkHttp 實現 WebSocket 通信感興趣,可以看看之前的文章《OkHttp實現WebSocket,包括鑑權和長連接保活及其原理》。
HTML5 擁有許多引人注目的新特性,WebSocket 就是其中之一。
WebSocket 一向有着 "Web 的 TCP" 之稱。通常 WebSocket 都是用於 Web 的,用於構建實時的 Web 應用。它可以在瀏覽器和服務器之間提供一個基於 TCP 連接的雙向通道。
WebSocket 協議本質上是一個基於 TCP 的協議。爲了建立一個 WebSocket 連接,瀏覽器首先要向服務器發起一個 HTTP 請求,這個請求和通常的 HTTP 請求不同,包含了一些附加頭信息,其中附加頭信息 "Upgrade: WebSocket" 表明這是一個申請協議升級的 HTTP 請求,服務器端解析這些附加的頭信息然後產生應答信息,返回給客戶端,客戶端和服務器端的 WebSocket 連接就建立起來了,雙方就可以通過這個連接通道自由的傳遞信息,並且這個連接會持續存在直到客戶端或者服務器端的某一方主動的關閉連接。
Websocket 同樣可以用於移動端。儘管移動端 Android/iOS 的本地應用可以直接通過 Socket 與服務器建立連接,並定義自己的協議來解決 Web 中實時應用創建困難的問題,但 WebSocket 服務通常複用 Web 的 80 端口,且可以比較方便的基於 Web 服務器來實現,因而對於某些端口容易被封的網絡環境而言,WebSocket 就變得非常有意義。
OkHttp 是在 2016 年 6 月 10 日發佈的 3.4.1 版中添加的對 WebSocket 的支持的。本文通過分析 OkHttp-3.5.0 的 WebSocket 實現來學習一下這個協議。
一、OkHttp API 使用
在開始分析 WebSocket 的實現之前,我們先來看一下 OkHttp 的 WebSocket API 怎麼用。
示例代碼如下:
public class WebsocketClient {
private static final int NORMAL_CLOSURE_STATUS = 1000;
private static OkHttpClient sClient;
private static WebSocket sWebSocket;
public static synchronized void startRequest() {
if (sClient == null) {
sClient = new OkHttpClient();
}
if (sWebSocket == null) {
Request request = new Request.Builder().url("ws://echo.websocket.org").build();
EchoWebSocketListener listener = new EchoWebSocketListener();
sWebSocket = sClient.newWebSocket(request, listener);
}
}
private static void sendMessage(WebSocket webSocket) {
webSocket.send("Knock, knock!");
webSocket.send("Hello!");
webSocket.send(ByteString.decodeHex("deadbeef"));
}
public static void sendMessage() {
WebSocket webSocket;
synchronized (WebsocketClient.class) {
webSocket = sWebSocket;
}
if (webSocket != null) {
sendMessage(webSocket);
}
}
public static synchronized void closeWebSocket() {
if (sWebSocket != null) {
sWebSocket.close(NORMAL_CLOSURE_STATUS, "Goodbye!");
sWebSocket = null;
}
}
public static synchronized void destroy() {
if (sClient != null) {
sClient.dispatcher().executorService().shutdown();
sClient = null;
}
}
private static void resetWebSocket() {
synchronized (WebsocketClient.class) {
sWebSocket = null;
}
}
public static class EchoWebSocketListener extends WebSocketListener {
private static final String TAG = "EchoWebSocketListener";
@Override
public void onOpen(WebSocket webSocket, Response response) {
sendMessage(webSocket);
}
@Override
public void onMessage(WebSocket webSocket, String text) {
Log.i(TAG, "Receiving: " + text);
}
@Override
public void onMessage(WebSocket webSocket, ByteString bytes) {
Log.i(TAG, "Receiving: " + bytes.hex());
}
@Override
public void onClosing(WebSocket webSocket, int code, String reason) {
webSocket.close(NORMAL_CLOSURE_STATUS, null);
Log.i(TAG, "Closing: " + code + " " + reason);
resetWebSocket();
}
@Override
public void onClosed(WebSocket webSocket, int code, String reason) {
Log.i(TAG, "Closed: " + code + " " + reason);
}
@Override
public void onFailure(WebSocket webSocket, Throwable t, Response response) {
t.printStackTrace();
resetWebSocket();
}
}
}
這個過程與發送 HTTP 請求的過程有許多相似之處,它們都需要創建 OkHttpClient 和 Request。然而它們不同的地方更多:
WebSocket 請求通過 WebSocketListener 來接收連接的狀態和活動,而 HTTP 請求則通過 Callback。同時請求的 URL 的 scheme 是 "ws" 或者是 "wss" (TLS 之上的 WebSocket),而不是 HTTP 的 "http" 和 "https";
HTTP 請求的連接建立及執行需要基於 Request 和回調創建 Call,並調用 Call 的方法手動進行;而對於 WebSocket 請求,則在基於 Request 和回調創建 WebSocket 的時候,OkHttp 會自動發起連接建立的過程;
這也是 WebSocket 與 HTTP 最大的不同。對於 WebSocket,我們可以保存 WebSocket 對象,並在後續多次通過該對象向服務器發送數據;
通過回調可以獲得更多 WebSocket 的狀態變化。在連接建立、收到服務器發送回來的消息、服務器要關閉連接,以及出現 error 時,都能得到通知。不像 HTTP 請求那樣,只在最後得到一個請求成功或者失敗的結果;
後兩點正是 WebSocket 全雙工連接的體現。
二、OkHttp 的 WebSocket 實現
接着我們來看 OkHttp 的 WebSocket 實現。
WebSocket 包含兩個部分,分別是握手和數據傳輸,數據傳輸又包括數據的發送,數據的接收,連接的保活,以及連接的關閉等,我們將分別分析這些過程。
2.1 連接握手
創建 WebSocket 的過程如下:
public class OkHttpClient implements Cloneable, Call.Factory, WebSocket.Factory {
// ...
@Override public WebSocket newWebSocket(Request request, WebSocketListener listener) {
RealWebSocket webSocket = new RealWebSocket(request, listener, new SecureRandom());
webSocket.connect(this);
return webSocket;
}
在這裏會創建一個 RealWebSocket
對象,然後執行其 connect()
方法建立連接。RealWebSocket
對象的創建過程如下:
public final class RealWebSocket implements WebSocket, WebSocketReader.FrameCallback {
// ...
public RealWebSocket(Request request, WebSocketListener listener, Random random) {
if (!"GET".equals(request.method())) {
throw new IllegalArgumentException("Request must be GET: " + request.method());
}
this.originalRequest = request;
this.listener = listener;
this.random = random;
byte[] nonce = new byte[16];
random.nextBytes(nonce);
this.key = ByteString.of(nonce).base64();
this.writerRunnable = new Runnable() {
@Override public void run() {
try {
while (writeOneFrame()) {
}
} catch (IOException e) {
failWebSocket(e, null);
}
}
};
}
這裏最主要的是初始化了 key,以備後續連接建立及握手之用。Key 是一個 16 字節長的隨機數經過 Base64 編碼得到的。此外還初始化了 writerRunnable
等。
連接建立及握手過程如下:
public final class RealWebSocket implements WebSocket, WebSocketReader.FrameCallback {
// ...
public void connect(OkHttpClient client) {
client = client.newBuilder()
.protocols(ONLY_HTTP1)
.build();
final int pingIntervalMillis = client.pingIntervalMillis();
final Request request = originalRequest.newBuilder()
.header("Upgrade", "websocket")
.header("Connection", "Upgrade")
.header("Sec-WebSocket-Key", key)
.header("Sec-WebSocket-Version", "13")
.build();
call = Internal.instance.newWebSocketCall(client, request);
call.enqueue(new Callback() {
@Override public void onResponse(Call call, Response response) {
try {
checkResponse(response);
} catch (ProtocolException e) {
failWebSocket(e, response);
closeQuietly(response);
return;
}
// Promote the HTTP streams into web socket streams.
StreamAllocation streamAllocation = Internal.instance.streamAllocation(call);
streamAllocation.noNewStreams(); // Prevent connection pooling!
Streams streams = streamAllocation.connection().newWebSocketStreams(streamAllocation);
// Process all web socket messages.
try {
listener.onOpen(RealWebSocket.this, response);
String name = "OkHttp WebSocket " + request.url().redact();
initReaderAndWriter(name, pingIntervalMillis, streams);
streamAllocation.connection().socket().setSoTimeout(0);
loopReader();
} catch (Exception e) {
failWebSocket(e, null);
}
}
@Override public void onFailure(Call call, IOException e) {
failWebSocket(e, null);
}
});
}
連接建立及握手的過程主要是向服務器發送一個 HTTP 請求。這個 HTTP 請求的特別之處在於,它包含了如下的一些 Headers:
Upgrade: WebSocket
Connection: Upgrade
Sec-WebSocket-Key: 7wgaspE0Tl7/66o4Dov2kw==
Sec-WebSocket-Version: 13
其中 Upgrade
和 Connection
header 向服務器表明,請求的目的就是要將客戶端和服務器端的通訊協議從 HTTP 協議升級到 WebSocket 協議,同時在請求處理完成之後,連接不要斷開。
Sec-WebSocket-Key
header 值正是我們前面看到的 key,它是 WebSocket 客戶端發送的一個 base64 編碼的密文,要求服務端必須返回一個對應加密的 "Sec-WebSocket-Accept" 應答,否則客戶端會拋出 "Error during WebSocket handshake" 錯誤,並關閉連接。
來自於 HTTP 服務器的響應到達的時候,即是連接建立大功告成的時候,也就是熱豆腐熟了的時候。
然而,響應到達時,儘管連接已經建立,還要爲數據的收發做一些準備。這些準備中的第一步就是檢查 HTTP 響應:
public final class RealWebSocket implements WebSocket, WebSocketReader.FrameCallback {
// ...
void checkResponse(Response response) throws ProtocolException {
if (response.code() != 101) {
throw new ProtocolException("Expected HTTP 101 response but was '"
+ response.code() + " " + response.message() + "'");
}
String headerConnection = response.header("Connection");
if (!"Upgrade".equalsIgnoreCase(headerConnection)) {
throw new ProtocolException("Expected 'Connection' header value 'Upgrade' but was '"
+ headerConnection + "'");
}
String headerUpgrade = response.header("Upgrade");
if (!"websocket".equalsIgnoreCase(headerUpgrade)) {
throw new ProtocolException(
"Expected 'Upgrade' header value 'websocket' but was '" + headerUpgrade + "'");
}
String headerAccept = response.header("Sec-WebSocket-Accept");
String acceptExpected = ByteString.encodeUtf8(key + WebSocketProtocol.ACCEPT_MAGIC)
.sha1().base64();
if (!acceptExpected.equals(headerAccept)) {
throw new ProtocolException("Expected 'Sec-WebSocket-Accept' header value '"
+ acceptExpected + "' but was '" + headerAccept + "'");
}
}
// ...
public void failWebSocket(Exception e, Response response) {
Streams streamsToClose;
synchronized (this) {
if (failed) return; // Already failed.
failed = true;
streamsToClose = this.streams;
this.streams = null;
if (cancelFuture != null) cancelFuture.cancel(false);
if (executor != null) executor.shutdown();
}
try {
listener.onFailure(this, e, response);
} finally {
closeQuietly(streamsToClose);
}
}
根據 WebSocket 的協議,服務器端用如下響應,來表示接受建立 WebSocket 連接的請求:
響應碼是 101;
"Connection" header 的值爲 "Upgrade",以表明服務器並沒有在處理完請求之後把連接斷開;
"Upgrade" header 的值爲 "websocket",以表明服務器接受後面使用 WebSocket 來通信;
"Sec-WebSocket-Accept" header 的值爲,key + WebSocketProtocol.ACCEPT_MAGIC 做 SHA1 hash,然後做 base64 編碼,來做服務器接受連接的驗證。關於這部分的設計的詳細信息,可參考 《WebSocket 協議規範》。
爲數據收發做準備的第二步是,初始化用於輸入輸出的 Source 和 Sink。Source 和 Sink 創建於之前發送 HTTP 請求的時候。這裏會阻止在這個連接上再創建新的流。
public final class RealConnection extends Http2Connection.Listener implements Connection {
// ...
public RealWebSocket.Streams newWebSocketStreams(final StreamAllocation streamAllocation) {
return new RealWebSocket.Streams(true, source, sink) {
@Override public void close() throws IOException {
streamAllocation.streamFinished(true, streamAllocation.codec());
}
};
}
Streams 是一個 BufferedSource 和 BufferedSink 的 holder:
public final class RealWebSocket implements WebSocket, WebSocketReader.FrameCallback {
// ...
public abstract static class Streams implements Closeable {
public final boolean client;
public final BufferedSource source;
public final BufferedSink sink;
public Streams(boolean client, BufferedSource source, BufferedSink sink) {
this.client = client;
this.source = source;
this.sink = sink;
}
}
第三步是調用回調 onOpen()
。
第四步是初始化 Reader 和 Writer:
public final class RealWebSocket implements WebSocket, WebSocketReader.FrameCallback {
// ...
public void initReaderAndWriter(
String name, long pingIntervalMillis, Streams streams) throws IOException {
synchronized (this) {
this.streams = streams;
this.writer = new WebSocketWriter(streams.client, streams.sink, random);
this.executor = new ScheduledThreadPoolExecutor(1, Util.threadFactory(name, false));
if (pingIntervalMillis != 0) {
executor.scheduleAtFixedRate(
new PingRunnable(), pingIntervalMillis, pingIntervalMillis, MILLISECONDS);
}
if (!messageAndCloseQueue.isEmpty()) {
runWriter(); // Send messages that were enqueued before we were connected.
}
}
reader = new WebSocketReader(streams.client, streams.source, this);
}
OkHttp 使用 WebSocketReader
和 WebSocketWriter
來處理數據的收發。在發送數據時將數據組織成幀,在接收數據時則進行反向擦除,同時處理 WebSocket 的控制消息。
WebSocket 的所有數據發送動作,都會在單線程線程池的線程中,通過 WebSocketWriter 執行。在這裏會創建 ScheduledThreadPoolExecutor 用於跑數據的發送操作。
WebSocket 協議中主要會傳輸兩種類型的幀:
控制幀,主要是用於連接保活的 Ping 幀等;
數據載荷幀。在這裏會根據用戶的配置,調度 Ping 幀週期性地發送。我們在調用 WebSocket 的接口發送數據時,數據並不是同步發送的,而是被放在了一個消息隊列中。發送消息的 Runnable 從消息隊列中讀取數據發送。其中會檢查消息隊列中是否有數據,如果有的話,會調度發送消息的 Runnable 執行。
第五步是配置 socket 的超時時間爲 0,也就是阻塞 IO。
第六步執行 loopReader()
。這實際上是進入了消息讀取循環了,也就是數據接收的邏輯了。
2.2 數據發送
我們可以通過 WebSocket 接口的 send(String text)
和 send(ByteString bytes)
分別發送文本的和二進制格式的消息。
public final class RealWebSocket implements WebSocket, WebSocketReader.FrameCallback {
// ...
@Override public boolean send(String text) {
if (text == null) throw new NullPointerException("text == null");
return send(ByteString.encodeUtf8(text), OPCODE_TEXT);
}
@Override public boolean send(ByteString bytes) {
if (bytes == null) throw new NullPointerException("bytes == null");
return send(bytes, OPCODE_BINARY);
}
private synchronized boolean send(ByteString data, int formatOpcode) {
// Don't send new frames after we've failed or enqueued a close frame.
if (failed || enqueuedClose) return false;
// If this frame overflows the buffer, reject it and close the web socket.
if (queueSize + data.size() > MAX_QUEUE_SIZE) {
close(CLOSE_CLIENT_GOING_AWAY, null);
return false;
}
// Enqueue the message frame.
queueSize += data.size();
messageAndCloseQueue.add(new Message(formatOpcode, data));
runWriter();
return true;
}
// ...
private void runWriter() {
assert (Thread.holdsLock(this));
if (executor != null) {
executor.execute(writerRunnable);
}
}
可以看到我們調用發送數據的接口時,做的事情主要是將數據格式化,構造消息,放進一個消息隊列,然後調度 writerRunnable 執行。
此外,值得注意的是,當消息隊列中的未發送數據,超出最大大小限制,WebSocket 連接會被直接關閉。對於發送失敗過或被關閉了的 WebSocket,將無法再發送信息。
在 writerRunnable
中會循環調用 writeOneFrame()
逐幀發送數據,直到數據發完,或發送失敗。在 WebSocket 協議中,客戶端需要發送 4 種類型的幀:
PING 幀
PONG 幀
CLOSE 幀
MESSAGE 幀
PING 幀用於連接保活,它的發送是在 PingRunnable
中執行的,在初始化 Reader 和 Writer 的時候,就會根據設置調度執行或不執行。
除 PING 幀外的其它三種幀,都在 writeOneFrame()
中發送。PONG 幀是對服務器發過來的 PING 幀的響應,同樣用於保活連接。
後面我們在分析連接的保活時會更詳細的分析 PING 和 PONG 這兩種幀。CLOSE 幀用於關閉連接,稍後我們在分析連接關閉過程時再來詳細地分析。
這裏我們主要關注用戶數據發送的部分。PONG 幀具有最高的發送優先級。在沒有 PONG 幀需要發送時,writeOneFrame()
從消息隊列中取出一條消息,如果消息不是 CLOSE 幀,則主要通過如下的過程進行發送:
public final class RealWebSocket implements WebSocket, WebSocketReader.FrameCallback {
// ...
boolean writeOneFrame() throws IOException {
WebSocketWriter writer;
ByteString pong;
Object messageOrClose = null;
int receivedCloseCode = -1;
String receivedCloseReason = null;
Streams streamsToClose = null;
synchronized (RealWebSocket.this) {
if (failed) {
return false; // Failed web socket.
}
writer = this.writer;
pong = pongQueue.poll();
if (pong == null) {
messageOrClose = messageAndCloseQueue.poll();
// ...
} else if (messageOrClose instanceof Message) {
ByteString data = ((Message) messageOrClose).data;
BufferedSink sink = Okio.buffer(writer.newMessageSink(
((Message) messageOrClose).formatOpcode, data.size()));
sink.write(data);
sink.close();
synchronized (this) {
queueSize -= data.size();
}
} else if (messageOrClose instanceof Close) {
數據發送的過程可以總結如下:
創建一個 BufferedSink 用於數據發送;
將數據寫入前面創建的 BufferedSink 中;
關閉 BufferedSink;
更新 queueSize 以正確地指示未發送數據的長度;
這裏面的玄機主要在創建的 BufferedSink。創建的 Sink 是一個 FrameSink
:
static void toggleMask(byte[] buffer, long byteCount, byte[] key, long frameBytesRead) {
int keyLength = key.length;
for (int i = 0; i < byteCount; i++, frameBytesRead++) {
int keyIndex = (int) (frameBytesRead % keyLength);
buffer[i] = (byte) (buffer[i] ^ key[keyIndex]);
}
}
// ...
Sink newMessageSink(int formatOpcode, long contentLength) {
if (activeWriter) {
throw new IllegalStateException("Another message writer is active. Did you call close()?");
}
activeWriter = true;
// Reset FrameSink state for a new writer.
frameSink.formatOpcode = formatOpcode;
frameSink.contentLength = contentLength;
frameSink.isFirstFrame = true;
frameSink.closed = false;
return frameSink;
}
void writeMessageFrameSynchronized(int formatOpcode, long byteCount, boolean isFirstFrame,
boolean isFinal) throws IOException {
assert Thread.holdsLock(this);
if (writerClosed) throw new IOException("closed");
int b0 = isFirstFrame ? formatOpcode : OPCODE_CONTINUATION;
if (isFinal) {
b0 |= B0_FLAG_FIN;
}
sink.writeByte(b0);
int b1 = 0;
if (isClient) {
b1 |= B1_FLAG_MASK;
}
if (byteCount <= PAYLOAD_BYTE_MAX) {
b1 |= (int) byteCount;
sink.writeByte(b1);
} else if (byteCount <= PAYLOAD_SHORT_MAX) {
b1 |= PAYLOAD_SHORT;
sink.writeByte(b1);
sink.writeShort((int) byteCount);
} else {
b1 |= PAYLOAD_LONG;
sink.writeByte(b1);
sink.writeLong(byteCount);
}
if (isClient) {
random.nextBytes(maskKey);
sink.write(maskKey);
for (long written = 0; written < byteCount; ) {
int toRead = (int) Math.min(byteCount, maskBuffer.length);
int read = buffer.read(maskBuffer, 0, toRead);
if (read == -1) throw new AssertionError();
toggleMask(maskBuffer, read, maskKey, written);
sink.write(maskBuffer, 0, read);
written += read;
}
} else {
sink.write(buffer, byteCount);
}
sink.emit();
}
final class FrameSink implements Sink {
int formatOpcode;
long contentLength;
boolean isFirstFrame;
boolean closed;
@Override public void write(Buffer source, long byteCount) throws IOException {
if (closed) throw new IOException("closed");
buffer.write(source, byteCount);
// Determine if this is a buffered write which we can defer until close() flushes.
boolean deferWrite = isFirstFrame
&& contentLength != -1
&& buffer.size() > contentLength - 8192 /* segment size */;
long emitCount = buffer.completeSegmentByteCount();
if (emitCount > 0 && !deferWrite) {
synchronized (WebSocketWriter.this) {
writeMessageFrameSynchronized(formatOpcode, emitCount, isFirstFrame, false /* final */);
}
isFirstFrame = false;
}
}
@Override public void flush() throws IOException {
if (closed) throw new IOException("closed");
synchronized (WebSocketWriter.this) {
writeMessageFrameSynchronized(formatOpcode, buffer.size(), isFirstFrame, false /* final */);
}
isFirstFrame = false;
}
@Override public Timeout timeout() {
return sink.timeout();
}
@SuppressWarnings("PointlessBitwiseExpression")
@Override public void close() throws IOException {
if (closed) throw new IOException("closed");
synchronized (WebSocketWriter.this) {
writeMessageFrameSynchronized(formatOpcode, buffer.size(), isFirstFrame, true /* final */);
}
closed = true;
activeWriter = false;
}
}
FrameSink
的 write()
會先將數據寫入一個 Buffer 中,然後再從這個 Buffer 中讀取數據來發送。如果是第一次發送數據,同時剩餘要發送的數據小於 8192 字節時,會延遲執行實際的數據發送,等 close()
時刷新。
根據 RealWebSocket
的 writeOneFrame()
的邏輯,在 write()
時,總是寫入整個消息的所有數據,因而在 FrameSink
的 write()
中總是不會發送數據的。
writeMessageFrameSynchronized()
將用戶數據格式化併發送出去。規範中定義的數據格式如下:
基本結構爲:
第一個字節是 meta data 控制位,包括 4 位的操作碼,用於指明這是否是消息的最後一幀的 FIN 位及三個保留位;
第二個字節包括掩碼位,和載荷長度或載荷長度指示。只有載荷長度比較小,在 127 以內時,載荷長度纔會包含在這個字節中。否則這個字節中將包含載荷長度指示的位;
可選的載荷長度。載荷長度大於 127 時,幀中會專門有一些字節來描述載荷的長度。載荷長度具體佔用幾個字節,因載荷的實際長度而異;
可選的掩碼字節。客戶端發送的幀,設置掩碼指示位,幷包含四個字節的掩碼字節;
載荷數據。客戶端發送的數據,會將原始的數據與掩碼字節做異或之後再發送;
關於幀格式的更詳細信息,可以參考 《WebSocket Protocol 規範》。
2.3 數據的接收
如我們前面看到的, 在握手的 HTTP 請求返回之後,會在 HTTP 請求的回調裏,啓動消息讀取循環 loopReader()
:
public final class RealWebSocket implements WebSocket, WebSocketReader.FrameCallback {
// ...
/** Receive frames until there are no more. Invoked only by the reader thread. */
public void loopReader() throws IOException {
while (receivedCloseCode == -1) {
// This method call results in one or more onRead* methods being called on this thread.
reader.processNextFrame();
}
}
在這個循環中,不斷通過 WebSocketReader
的 processNextFrame()
讀取消息,直到收到了關閉連接的消息。
final class WebSocketReader {
public interface FrameCallback {
void onReadMessage(String text) throws IOException;
void onReadMessage(ByteString bytes) throws IOException;
void onReadPing(ByteString buffer);
void onReadPong(ByteString buffer);
void onReadClose(int code, String reason);
}
// ...
void processNextFrame() throws IOException {
readHeader();
if (isControlFrame) {
readControlFrame();
} else {
readMessageFrame();
}
}
private void readHeader() throws IOException {
if (closed) throw new IOException("closed");
// Disable the timeout to read the first byte of a new frame.
int b0;
long timeoutBefore = source.timeout().timeoutNanos();
source.timeout().clearTimeout();
try {
b0 = source.readByte() & 0xff;
} finally {
source.timeout().timeout(timeoutBefore, TimeUnit.NANOSECONDS);
}
opcode = b0 & B0_MASK_OPCODE;
isFinalFrame = (b0 & B0_FLAG_FIN) != 0;
isControlFrame = (b0 & OPCODE_FLAG_CONTROL) != 0;
// Control frames must be final frames (cannot contain continuations).
if (isControlFrame && !isFinalFrame) {
throw new ProtocolException("Control frames must be final.");
}
boolean reservedFlag1 = (b0 & B0_FLAG_RSV1) != 0;
boolean reservedFlag2 = (b0 & B0_FLAG_RSV2) != 0;
boolean reservedFlag3 = (b0 & B0_FLAG_RSV3) != 0;
if (reservedFlag1 || reservedFlag2 || reservedFlag3) {
// Reserved flags are for extensions which we currently do not support.
throw new ProtocolException("Reserved flags are unsupported.");
}
int b1 = source.readByte() & 0xff;
isMasked = (b1 & B1_FLAG_MASK) != 0;
if (isMasked == isClient) {
// Masked payloads must be read on the server. Unmasked payloads must be read on the client.
throw new ProtocolException(isClient
? "Server-sent frames must not be masked."
: "Client-sent frames must be masked.");
}
// Get frame length, optionally reading from follow-up bytes if indicated by special values.
frameLength = b1 & B1_MASK_LENGTH;
if (frameLength == PAYLOAD_SHORT) {
frameLength = source.readShort() & 0xffffL; // Value is unsigned.
} else if (frameLength == PAYLOAD_LONG) {
frameLength = source.readLong();
if (frameLength < 0) {
throw new ProtocolException(
"Frame length 0x" + Long.toHexString(frameLength) + " > 0x7FFFFFFFFFFFFFFF");
}
}
frameBytesRead = 0;
if (isControlFrame && frameLength > PAYLOAD_BYTE_MAX) {
throw new ProtocolException("Control frame must be less than " + PAYLOAD_BYTE_MAX + "B.");
}
if (isMasked) {
// Read the masking key as bytes so that they can be used directly for unmasking.
source.readFully(maskKey);
}
}
processNextFrame()
先讀取 Header 的兩個字節,然後根據 Header 的信息,讀取數據內容。
在讀取 Header 時,讀的第一個字節是同步的不計超時時間的。WebSocketReader
從 Header 中,獲取到這個幀是不是消息的最後一幀,消息的類型,是否有掩碼字節,保留位,幀的長度,以及掩碼字節等信息。
WebSocket 通過掩碼位和掩碼字節來區分數據是從客戶端發送給服務器的,還是服務器發送給客戶端的。這裏會根據協議,對這些信息進行有效性一致性檢驗,若不一致則會拋出 ProtocolException
。
WebSocketReader
同步讀取時的調用棧如下:
通過幀的 Header 確定了是數據幀,則會執行 readMessageFrame()
讀取消息幀:
final class WebSocketReader {
// ...
private void readMessageFrame() throws IOException {
int opcode = this.opcode;
if (opcode != OPCODE_TEXT && opcode != OPCODE_BINARY) {
throw new ProtocolException("Unknown opcode: " + toHexString(opcode));
}
Buffer message = new Buffer();
readMessage(message);
if (opcode == OPCODE_TEXT) {
frameCallback.onReadMessage(message.readUtf8());
} else {
frameCallback.onReadMessage(message.readByteString());
}
}
/** Read headers and process any control frames until we reach a non-control frame. */
void readUntilNonControlFrame() throws IOException {
while (!closed) {
readHeader();
if (!isControlFrame) {
break;
}
readControlFrame();
}
}
/**
* Reads a message body into across one or more frames. Control frames that occur between
* fragments will be processed. If the message payload is masked this will unmask as it's being
* processed.
*/
private void readMessage(Buffer sink) throws IOException {
while (true) {
if (closed) throw new IOException("closed");
if (frameBytesRead == frameLength) {
if (isFinalFrame) return; // We are exhausted and have no continuations.
readUntilNonControlFrame();
if (opcode != OPCODE_CONTINUATION) {
throw new ProtocolException("Expected continuation opcode. Got: " + toHexString(opcode));
}
if (isFinalFrame && frameLength == 0) {
return; // Fast-path for empty final frame.
}
}
long toRead = frameLength - frameBytesRead;
long read;
if (isMasked) {
toRead = Math.min(toRead, maskBuffer.length);
read = source.read(maskBuffer, 0, (int) toRead);
if (read == -1) throw new EOFException();
toggleMask(maskBuffer, read, maskKey, frameBytesRead);
sink.write(maskBuffer, 0, (int) read);
} else {
read = source.read(sink, toRead);
if (read == -1) throw new EOFException();
}
frameBytesRead += read;
}
}
這個過程中,會讀取一條消息包含的所有數據幀。按照 WebSocket 的標準,包含用戶數據的消息數據幀可以和控制幀交替發送。
但消息之間的數據幀不可以。因而在這個過程中,若遇到了控制幀,則會先讀取控制幀進行處理,然後繼續讀取消息的數據幀,直到讀取了消息的所有數據幀。
掩碼位和掩碼字節,對於客戶端而言,發送的數據中包含這些東西,在接收的數據中不包含這些。對於服務器而言,則是在接收的數據中包含這些,發送的數據中不包含。OkHttp 既支持服務器開發,也支持客戶端開發,因而可以看到對於掩碼位和掩碼字節完整的處理。
在一個消息讀取完成之後,會通過回調 FrameCallback
將讀取的內容通知出去。
final class WebSocketReader {
// ...
WebSocketReader(boolean isClient, BufferedSource source, FrameCallback frameCallback) {
if (source == null) throw new NullPointerException("source == null");
if (frameCallback == null) throw new NullPointerException("frameCallback == null");
this.isClient = isClient;
this.source = source;
this.frameCallback = frameCallback;
}
這一事件會通知到 RealWebSocket
。
public final class RealWebSocket implements WebSocket, WebSocketReader.FrameCallback {
// ...
@Override public void onReadMessage(String text) throws IOException {
listener.onMessage(this, text);
}
@Override public void onReadMessage(ByteString bytes) throws IOException {
listener.onMessage(this, bytes);
}
在 RealWebSocket
中,這一事件又被通知到我們在應用程序中創建的回調 WebSocketListener
。
2.4 連接的保活
連接的保活通過 PING 幀和 PONG 幀來實現。如我們前面看到的,若用戶設置了 PING 幀的發送週期,在握手的 HTTP 請求返回時,消息讀取循環開始前會調度 PingRunnable
週期性的向服務器發送 PING 幀:
public final class RealWebSocket implements WebSocket, WebSocketReader.FrameCallback {
// ...
private final class PingRunnable implements Runnable {
PingRunnable() {
}
@Override public void run() {
writePingFrame();
}
}
void writePingFrame() {
WebSocketWriter writer;
synchronized (this) {
if (failed) return;
writer = this.writer;
}
try {
writer.writePing(ByteString.EMPTY);
} catch (IOException e) {
failWebSocket(e, null);
}
}
在 PingRunnable
中,通過 WebSocketWriter
發送 PING 幀:
final class WebSocketWriter {
// ...
/** Send a ping with the supplied {@code payload}. */
void writePing(ByteString payload) throws IOException {
synchronized (this) {
writeControlFrameSynchronized(OPCODE_CONTROL_PING, payload);
}
}
// ...
private void writeControlFrameSynchronized(int opcode, ByteString payload) throws IOException {
assert Thread.holdsLock(this);
if (writerClosed) throw new IOException("closed");
int length = payload.size();
if (length > PAYLOAD_BYTE_MAX) {
throw new IllegalArgumentException(
"Payload size must be less than or equal to " + PAYLOAD_BYTE_MAX);
}
int b0 = B0_FLAG_FIN | opcode;
sink.writeByte(b0);
int b1 = length;
if (isClient) {
b1 |= B1_FLAG_MASK;
sink.writeByte(b1);
random.nextBytes(maskKey);
sink.write(maskKey);
byte[] bytes = payload.toByteArray();
toggleMask(bytes, bytes.length, maskKey, 0);
sink.write(bytes);
} else {
sink.writeByte(b1);
sink.write(payload);
}
sink.flush();
}
PING 幀是一個不包含載荷的控制幀。關於掩碼位和掩碼字節的設置,與消息的數據幀相同。即客戶端發送的幀,設置掩碼位,幀中包含掩碼字節。服務器發送的幀,不設置掩碼位,幀中不包含掩碼字節。
通過 WebSocket 通信的雙方,在收到對方發來的 PING 幀時,需要用 PONG 幀來回復。在 WebSocketReader
的 readControlFrame()
中可以看到這一點:
final class WebSocketReader {
// ...
private void readControlFrame() throws IOException {
Buffer buffer = new Buffer();
if (frameBytesRead < frameLength) {
if (isClient) {
source.readFully(buffer, frameLength);
} else {
while (frameBytesRead < frameLength) {
int toRead = (int) Math.min(frameLength - frameBytesRead, maskBuffer.length);
int read = source.read(maskBuffer, 0, toRead);
if (read == -1) throw new EOFException();
toggleMask(maskBuffer, read, maskKey, frameBytesRead);
buffer.write(maskBuffer, 0, read);
frameBytesRead += read;
}
}
}
switch (opcode) {
case OPCODE_CONTROL_PING:
frameCallback.onReadPing(buffer.readByteString());
break;
case OPCODE_CONTROL_PONG:
frameCallback.onReadPong(buffer.readByteString());
break;
PING 幀和 PONG 幀都不帶載荷,控制幀讀寫時對於載荷長度的處理,都是爲 CLOSE 幀做的。因而針對 PING 幀和 PONG 幀,除了 Header 外, readControlFrame()
實際上無需再讀取任何數據,但它會將這些事件通知出去:
public final class RealWebSocket implements WebSocket, WebSocketReader.FrameCallback {
// ...
@Override public synchronized void onReadPing(ByteString payload) {
// Don't respond to pings after we've failed or sent the close frame.
if (failed || (enqueuedClose && messageAndCloseQueue.isEmpty())) return;
pongQueue.add(payload);
runWriter();
pingCount++;
}
@Override public synchronized void onReadPong(ByteString buffer) {
// This API doesn't expose pings.
pongCount++;
}
可見在收到 PING 幀的時候,總是會發一個 PONG 幀出去,且通常其沒有載荷數據。在收到一個 PONG 幀時,則通常只是記錄一下,然後什麼也不做。如我們前面所見,PONG 幀在 writerRunnable
中被髮送出去:
public final class RealWebSocket implements WebSocket, WebSocketReader.FrameCallback {
// ...
if (pong != null) {
writer.writePong(pong);
} else if (messageOrClose instanceof Message) {
PONG 幀的發送與 PING 幀的非常相似:
final class WebSocketWriter {
// ...
/** Send a pong with the supplied {@code payload}. */
void writePong(ByteString payload) throws IOException {
synchronized (this) {
writeControlFrameSynchronized(OPCODE_CONTROL_PONG, payload);
}
}
2.5 連接的關閉
連接的關閉,與數據發送的過程頗有幾分相似之處。通過 WebSocket
接口的 close(int code, String reason)
我們可以關閉一個 WebSocket 連接:
public final class RealWebSocket implements WebSocket, WebSocketReader.FrameCallback {
// ...
@Override public boolean close(int code, String reason) {
return close(code, reason, CANCEL_AFTER_CLOSE_MILLIS);
}
synchronized boolean close(int code, String reason, long cancelAfterCloseMillis) {
validateCloseCode(code);
ByteString reasonBytes = null;
if (reason != null) {
reasonBytes = ByteString.encodeUtf8(reason);
if (reasonBytes.size() > CLOSE_MESSAGE_MAX) {
throw new IllegalArgumentException("reason.size() > " + CLOSE_MESSAGE_MAX + ": " + reason);
}
}
if (failed || enqueuedClose) return false;
// Immediately prevent further frames from being enqueued.
enqueuedClose = true;
// Enqueue the close frame.
messageAndCloseQueue.add(new Close(code, reasonBytes, cancelAfterCloseMillis));
runWriter();
return true;
}
在執行關閉連接動作前,會先檢查一下 close code 的有效性在合法範圍內。關於不同 close code 的詳細說明,可以參考 《WebSocket 協議規範》。
檢查完了之後,會構造一個 Close 消息放入發送消息隊列,並調度 writerRunnable
執行。Close 消息可以帶有不超出 123 字節的字符串,以作爲 Close message,來說明連接關閉的原因。
連接的關閉分爲主動關閉和被動關閉。客戶端先向服務器發送一個 CLOSE 幀,然後服務器恢復一個 CLOSE 幀,對於客戶端而言,這個過程爲主動關閉。反之則爲對客戶端而言則爲被動關閉。
在 writerRunnable
執行的 writeOneFrame()
實際發送 CLOSE 幀:
public final class RealWebSocket implements WebSocket, WebSocketReader.FrameCallback {
// ...
messageOrClose = messageAndCloseQueue.poll();
if (messageOrClose instanceof Close) {
receivedCloseCode = this.receivedCloseCode;
receivedCloseReason = this.receivedCloseReason;
if (receivedCloseCode != -1) {
streamsToClose = this.streams;
this.streams = null;
this.executor.shutdown();
} else {
// When we request a graceful close also schedule a cancel of the websocket.
cancelFuture = executor.schedule(new CancelRunnable(),
((Close) messageOrClose).cancelAfterCloseMillis, MILLISECONDS);
}
} else if (messageOrClose == null) {
return false; // The queue is exhausted.
}
}
// ...
} else if (messageOrClose instanceof Close) {
Close close = (Close) messageOrClose;
writer.writeClose(close.code, close.reason);
// We closed the writer: now both reader and writer are closed.
if (streamsToClose != null) {
listener.onClosed(this, receivedCloseCode, receivedCloseReason);
}
} else {
發送 CLOSE 幀也分爲主動關閉的發送還是被動關閉的發送。
對於被動關閉,在發送完 CLOSE 幀之後,連接被最終關閉,因而,發送 CLOSE 幀之前,這裏會停掉髮送消息用的 executor。而在發送之後,則會通過 onClosed()
通知用戶。
而對於主動關閉,則在發送前會調度 CancelRunnable
的執行,發送後不會通過 onClosed()
通知用戶。
final class WebSocketWriter {
// ...
void writeClose(int code, ByteString reason) throws IOException {
ByteString payload = ByteString.EMPTY;
if (code != 0 || reason != null) {
if (code != 0) {
validateCloseCode(code);
}
Buffer buffer = new Buffer();
buffer.writeShort(code);
if (reason != null) {
buffer.write(reason);
}
payload = buffer.readByteString();
}
synchronized (this) {
try {
writeControlFrameSynchronized(OPCODE_CONTROL_CLOSE, payload);
} finally {
writerClosed = true;
}
}
}
將 CLOSE 幀發送到網絡的過程與 PING 和 PONG 幀的頗爲相似,僅有的差別就是 CLOSE 幀有載荷。關於掩碼位和掩碼自己的規則,同樣適用於 CLOSE 幀的發送。
CLOSE 的讀取在 WebSocketReader
的 readControlFrame()
中:
final class WebSocketReader {
// ...
private void readControlFrame() throws IOException {
Buffer buffer = new Buffer();
if (frameBytesRead < frameLength) {
if (isClient) {
source.readFully(buffer, frameLength);
} else {
while (frameBytesRead < frameLength) {
int toRead = (int) Math.min(frameLength - frameBytesRead, maskBuffer.length);
int read = source.read(maskBuffer, 0, toRead);
if (read == -1) throw new EOFException();
toggleMask(maskBuffer, read, maskKey, frameBytesRead);
buffer.write(maskBuffer, 0, read);
frameBytesRead += read;
}
}
}
switch (opcode) {
// ...
case OPCODE_CONTROL_CLOSE:
int code = CLOSE_NO_STATUS_CODE;
String reason = "";
long bufferSize = buffer.size();
if (bufferSize == 1) {
throw new ProtocolException("Malformed close payload length of 1.");
} else if (bufferSize != 0) {
code = buffer.readShort();
reason = buffer.readUtf8();
String codeExceptionMessage = WebSocketProtocol.closeCodeExceptionMessage(code);
if (codeExceptionMessage != null) throw new ProtocolException(codeExceptionMessage);
}
frameCallback.onReadClose(code, reason);
closed = true;
break;
default:
throw new ProtocolException("Unknown control opcode: " + toHexString(opcode));
}
}
讀到 CLOSE 幀時,WebSocketReader
會將這一事件通知出去:
public final class RealWebSocket implements WebSocket, WebSocketReader.FrameCallback {
// ...
@Override public void onReadClose(int code, String reason) {
if (code == -1) throw new IllegalArgumentException();
Streams toClose = null;
synchronized (this) {
if (receivedCloseCode != -1) throw new IllegalStateException("already closed");
receivedCloseCode = code;
receivedCloseReason = reason;
if (enqueuedClose && messageAndCloseQueue.isEmpty()) {
toClose = this.streams;
this.streams = null;
if (cancelFuture != null) cancelFuture.cancel(false);
this.executor.shutdown();
}
}
try {
listener.onClosing(this, code, reason);
if (toClose != null) {
listener.onClosed(this, code, reason);
}
} finally {
closeQuietly(toClose);
}
}
對於收到的 CLOSE 幀處理同樣分爲主動關閉的情況和被動關閉的情況。與 CLOSE 發送時的情形正好相反,若是主動關閉,則在收到 CLOSE 幀之後,WebSocket 連接最終斷開,因而需要停掉 executor,被動關閉則暫時不需要。
收到 CLOSE 幀,總是會通過 onClosing()
將事件通知出去。
對於主動關閉的情形,最後還會通過 onClosed()
通知用戶,連接已經最終關閉。
關於 WebSocket 的 CLOSE 幀的更多說明,可以參考 WebSocket 協議規範。
三、WebSocket 連接的生命週期
總結一下 WebSocket 連接的生命週期:
連接通過一個 HTTP 請求握手並建立連接。WebSocket 連接可以理解爲是通過 HTTP 請求建立的普通 TCP 連接。
WebSocket 做了二進制分幀。WebSocket 連接中收發的數據以幀爲單位。主要有用於連接保活的控制幀 PING 和 PONG,用於用戶數據發送的 MESSAGE 幀,和用於關閉連接的控制幀 CLOSE。
連接建立之後,通過 PING 幀和 PONG 幀做連接保活。
一次 send 數據,被封爲一個消息,通過一個或多個 MESSAGE 幀進行發送。一個消息的幀和控制幀可以交叉發送,不同消息的幀之間不可以。
WebSocket 連接的兩端相互發送一個 CLOSE 幀以最終關閉連接。
關於 WebSocket 的詳細信息,可以參考《 WebSocket 協議規範》。
references:
WebSocket 協議規範:https://tools.ietf.org/html/rfc6455
WebSocket 實戰:https://www.ibm.com/developerworks/cn/java/j-lo-WebSocket/index.html
使用 HTML5 WebSocket 構建實時 Web 應用:https://www.ibm.com/developerworks/cn/web/1112_huangxa_websocket/index.html
WebSocket Client Example with OkHttp:http://howtoprogram.xyz/2016/12/24/websocket-client-example-okhttp/
熱文推薦:
公衆號後臺回覆成長『成長』,將會得到我準備的學習資料。