go - 强制关闭 SSH 客户端 session

Question

我编写了一个 SSH 客户端来连接到网络设备，一旦运行命令超过 25 秒，我就通过“选择”设置超时。我注意到一些设备有另一个 IOS，一旦超时被触发并导致 goroutinge 泄漏，它就无法通过 Close() 方法删除与它们的 SSH session 。我需要跟上客户端并断开 session 以准备好下一个命令。看起来 goroutine 在那个时候不会永远终止!你们有什么想法吗？

    go func() {
       r <- s.Run(cmd)
    }()

    select {
       case err := <-r:
         return err
       case <-time.After(time.Duration(timeout) * time.Second):
         s.Close()
         return fmt.Errorf("timeout after %d seconds", timeout)
    }

通过堆分析，我看到了以下内容:
2.77GB 99.44% 99.44% 2.77GB 99.44% bytes.makeSlice

     0     0% 99.44%     2.77GB 99.44%  bytes.(*Buffer).ReadFrom

     0     0% 99.44%     2.77GB 99.44%  golang.org/x/crypto/ssh.(*Session).start.func1

     0     0% 99.44%     2.77GB 99.44%  golang.org/x/crypto/ssh.(*Session).stdout.func1

     0     0% 99.44%     2.77GB 99.44%  io.Copy

     0     0% 99.44%     2.77GB 99.44%  io.copyBuffer

     0     0% 99.44%     2.78GB 99.93%  runtime.goexit

例程 ========================/usr/local/go/src/runtime/asm_amd64.s 中的 runtime.goexit

     0     2.78GB (flat, cum) 99.93% of Total

     .          .   1993:   RET

     .          .   1994:

     .          .   1995:// The top-most function running on a goroutine

     .          .   1996:// returns to goexit+PCQuantum.

     .          .   1997:TEXT runtime·goexit(SB),NOSPLIT,$0-0

     .     2.78GB   1998:   BYTE    $0x90   // NOP

     .          .   1999:   CALL    runtime·goexit1(SB) // does not return

     .          .   2000:   // traceback from goexit1 must hit code range of goexit

     .          .   2001:   BYTE    $0x90   // NOP

     .          .   2002:


     .          .   2003:TEXT runtime·prefetcht0(SB),NOSPLIT,$0-8

Answer 1

channel r阻止 Go 例程返回，因为它没有被清空。我已经编写了您代码的改编版本并插入了一个 WaitGroup 来演示问题:

func main() {
    var wg sync.WaitGroup // This is only added for demonstration purposes
    s := new(clientSession)

    r := make(chan error)

    go func(s *clientSession) {
        wg.Add(1)
        r <- s.Run()
        wg.Done() // Will only be called after s.Run() is able to return
    }(s)

    fmt.Println("Client has been opened")

    select {
    case err := <-r:
        fmt.Println(err)
    case <-time.After(1 * time.Second):
        s.Close()
        fmt.Println("Timed out, closing")
    }

    wg.Wait() // Waits until wg.Done() is called.

    fmt.Println("Main finished successfully")
}

Go Playground 似乎终止了程序，所以我创建了一个 gist带有完整的可运行代码。当我们运行 incorrect.go :

$ go run incorrect.go
Client has been opened
Timed out, closing
fatal error: all goroutines are asleep - deadlock!
....

那是因为我们的代码在 wg.Wait() 上死锁了线。这表明 wg.Done()在 Go 例程中从未达到过。

正如评论所指出的，缓冲 channel 可以在这里提供帮助。但前提是你不再关心错误，在调用 s.Close() 之后

r := make(chan error, 1)

buffered.go 运行正常，但错误丢失:

$ go run buffered.go
Client has been opened
Timed out, closing
Main finished successfully

另一种选择是精确地排空 channel 1 次:

select {
    case err := <-r:
        fmt.Println(err)
    case <-time.After(1 * time.Second):
        s.Close()
        fmt.Println("Timed out, closing")
        fmt.Println(<-r)
    }

或者通过包装 select在 for循环(无缓冲 channel ):

X:
    for {
        select {
        case err := <-r:
            fmt.Println(err)
            break X // because we are in main(). Normally `return err`
        case <-time.After(1 * time.Second):
            s.Close()
            fmt.Println("Timed out, closing")
        }
    }

当我们运行 drain.go 我们看到错误也打印出来:

$ go run incorrect.go
Client has been opened
Timed out, closing
Run() closed
Main finished successfully

在现实世界中，一个人会运行多个 Go 程序。所以你会想在 for 上使用一些计数器。循环或进一步利用 WaitGroup 功能。