|
| 1 | +# 源码分析golang http shutdown优雅退出的原理 |
| 2 | + |
| 3 | +我们知道在go 1.8.x后,golang在http里加入了shutdown方法,用来控制优雅退出。什么是优雅退出? 简单说就是不处理新请求,但是会处理正在进行的请求,把旧请求都处理完,也就是都response之后,那么就退出。 |
| 4 | + |
| 5 | +社区里不少http graceful动态重启,平滑重启的库,大多是基于http.shutdown做的。平滑启动的原理很简单,fork子进程,继承listen fd, 老进程优雅退出。以前写过文章专门讲述在golang里如何实现平滑重启 (graceful reload)。有兴趣的朋友可以翻翻。 |
| 6 | + |
| 7 | +## http shutdown 源码分析 |
| 8 | + |
| 9 | +先来看下 http shutdown 的主方法实现逻辑。用atomic来做退出标记的状态,然后关闭各种的资源,然后一直阻塞的等待无空闲连接,每 500ms 轮询一次。 |
| 10 | + |
| 11 | +```go |
| 12 | +// xiaorui.cc |
| 13 | +var shutdownPollInterval = 500 * time.Millisecond |
| 14 | + |
| 15 | +func (srv *Server) Shutdown(ctx context.Context) error { |
| 16 | + // 标记退出的状态 |
| 17 | + atomic.StoreInt32(&srv.inShutdown, 1) |
| 18 | + srv.mu.Lock() |
| 19 | + // 关闭listen fd,新连接无法建立。 |
| 20 | + lnerr := srv.closeListenersLocked() |
| 21 | + |
| 22 | + // 把server.go的done chan给close掉,通知等待的worekr退出 |
| 23 | + srv.closeDoneChanLocked() |
| 24 | + |
| 25 | + // 执行回调方法,我们可以注册shutdown的回调方法 |
| 26 | + for _, f := range srv.onShutdown { |
| 27 | + go f() |
| 28 | + } |
| 29 | + |
| 30 | + // 每500ms来检查下,是否没有空闲的连接了,或者监听上游传递的ctx上下文。 |
| 31 | + ticker := time.NewTicker(shutdownPollInterval) |
| 32 | + defer ticker.Stop() |
| 33 | + for { |
| 34 | + if srv.closeIdleConns() { |
| 35 | + return lnerr |
| 36 | + } |
| 37 | + select { |
| 38 | + case <-ctx.Done(): |
| 39 | + return ctx.Err() |
| 40 | + case <-ticker.C: |
| 41 | + } |
| 42 | + } |
| 43 | +} |
| 44 | +``` |
| 45 | + |
| 46 | +是否没有空闲的连接 |
| 47 | +遍历连接,当客户单的连接已空闲,则关闭连接,并在 activeConn 连接池中剔除该连接。 |
| 48 | + |
| 49 | +```go |
| 50 | +func (s *Server) closeIdleConns() bool { |
| 51 | + s.mu.Lock() |
| 52 | + defer s.mu.Unlock() |
| 53 | + quiescent := true |
| 54 | + for c := range s.activeConn { |
| 55 | + st, unixSec := c.getState() |
| 56 | + if st == StateNew && unixSec < time.Now().Unix()-5 { |
| 57 | + st = StateIdle |
| 58 | + } |
| 59 | + if st != StateIdle || unixSec == 0 { |
| 60 | + quiescent = false |
| 61 | + continue |
| 62 | + } |
| 63 | + c.rwc.Close() |
| 64 | + delete(s.activeConn, c) |
| 65 | + } |
| 66 | + return quiescent |
| 67 | +} |
| 68 | +``` |
| 69 | + |
| 70 | +关闭server.doneChan和监听的文件描述符 |
| 71 | + |
| 72 | +```go |
| 73 | +// xiaorui.cc |
| 74 | +// 关闭doen chan |
| 75 | +func (s *Server) closeDoneChanLocked() { |
| 76 | + ch := s.getDoneChanLocked() |
| 77 | + select { |
| 78 | + case <-ch: |
| 79 | + // Already closed. Don't close again. |
| 80 | + default: |
| 81 | + // Safe to close here. We're the only closer, guarded |
| 82 | + // by s.mu. |
| 83 | + close(ch) |
| 84 | + } |
| 85 | +} |
| 86 | + |
| 87 | +// 关闭监听的fd |
| 88 | +func (s *Server) closeListenersLocked() error { |
| 89 | + var err error |
| 90 | + for ln := range s.listeners { |
| 91 | + if cerr := (*ln).Close(); cerr != nil && err == nil { |
| 92 | + err = cerr |
| 93 | + } |
| 94 | + delete(s.listeners, ln) |
| 95 | + } |
| 96 | + return err |
| 97 | +} |
| 98 | + |
| 99 | +// 关闭连接 |
| 100 | +func (c *conn) Close() error { |
| 101 | + if !c.ok() { |
| 102 | + return syscall.EINVAL |
| 103 | + } |
| 104 | + err := c.fd.Close() |
| 105 | + if err != nil { |
| 106 | + err = &OpError{Op: "close", Net: c.fd.net, Source: c.fd.laddr, Addr: c.fd.raddr, Err: err} |
| 107 | + } |
| 108 | + return err |
| 109 | +} |
| 110 | +``` |
| 111 | + |
| 112 | +这么一系列的操作后,server.go的serv主监听方法也就退出了。 |
| 113 | + |
| 114 | +```go |
| 115 | +// xiaorui.cc |
| 116 | +func (srv *Server) Serve(l net.Listener) error { |
| 117 | + ... |
| 118 | + for { |
| 119 | + rw, e := l.Accept() |
| 120 | + if e != nil { |
| 121 | + select { |
| 122 | + // 退出 |
| 123 | + case <-srv.getDoneChan(): |
| 124 | + return ErrServerClosed |
| 125 | + default: |
| 126 | + } |
| 127 | + ... |
| 128 | + return e |
| 129 | + } |
| 130 | + tempDelay = 0 |
| 131 | + c := srv.newConn(rw) |
| 132 | + c.setState(c.rwc, StateNew) // before Serve can return |
| 133 | + go c.serve(ctx) |
| 134 | + } |
| 135 | +} |
| 136 | +``` |
| 137 | + |
| 138 | +那么如何保证用户在请求完成后,再关闭连接的? |
| 139 | + |
| 140 | +```go |
| 141 | +// xiaorui.cc |
| 142 | + |
| 143 | +func (s *Server) doKeepAlives() bool { |
| 144 | + return atomic.LoadInt32(&s.disableKeepAlives) == 0 && !s.shuttingDown() |
| 145 | +} |
| 146 | + |
| 147 | + |
| 148 | +// Serve a new connection. |
| 149 | +func (c *conn) serve(ctx context.Context) { |
| 150 | + defer func() { |
| 151 | + ... xiaorui.cc ... |
| 152 | + if !c.hijacked() { |
| 153 | + // 关闭连接,并且标记退出 |
| 154 | + c.close() |
| 155 | + c.setState(c.rwc, StateClosed) |
| 156 | + } |
| 157 | + }() |
| 158 | + ... |
| 159 | + ctx, cancelCtx := context.WithCancel(ctx) |
| 160 | + c.cancelCtx = cancelCtx |
| 161 | + defer cancelCtx() |
| 162 | + |
| 163 | + c.r = &connReader{conn: c} |
| 164 | + c.bufr = newBufioReader(c.r) |
| 165 | + c.bufw = newBufioWriterSize(checkConnErrorWriter{c}, 4<<10) |
| 166 | + |
| 167 | + for { |
| 168 | + // 接收请求 |
| 169 | + w, err := c.readRequest(ctx) |
| 170 | + if c.r.remain != c.server.initialReadLimitSize() { |
| 171 | + c.setState(c.rwc, StateActive) |
| 172 | + } |
| 173 | + ... |
| 174 | + ... |
| 175 | + // 匹配路由及回调处理方法 |
| 176 | + serverHandler{c.server}.ServeHTTP(w, w.req) |
| 177 | + w.cancelCtx() |
| 178 | + if c.hijacked() { |
| 179 | + return |
| 180 | + } |
| 181 | + ... |
| 182 | + // 判断是否在shutdown mode, 选择退出 |
| 183 | + if !w.conn.server.doKeepAlives() { |
| 184 | + return |
| 185 | + } |
| 186 | + } |
| 187 | + ... |
| 188 | +``` |
| 189 | +
|
| 190 | +## 总结: |
| 191 | +
|
| 192 | +总觉得golang net/http的代码写得有点乱,应该能写得更好。我也看过不少golang标准库的源代码,最让我头疼的就是net/http了。😅 |
0 commit comments