golang请求头并发读写

Question

去环境:

GOARCH="amd64"

GOBIN="/root/"

GOEXE=""

GOHOSTARCH="amd64"

GOHOSTOS="linux"

GOOS="linux"

GOPATH="/data/workspace/kubernetes"

GORACE=""

GOROOT="/usr/local/go"

GOTOOLDIR="/usr/local/go/pkg/tool/linux_amd64"

GO15VENDOREXPERIMENT="1"

CC="gcc"

GOGCCFLAGS="-fPIC -m64 -pthread -fmessage-length=0"

CXX="g++"

CGO_ENABLED="1"

Go 版本:

go version go1.6.3 linux/amd64

这个问题发生在高负载的“性能测试环境”kube-apiserver 上。 kube-apiserver 崩溃并退出:

fatal error: concurrent map read and map write

goroutine 77930636 [running]:
runtime.throw(0x2f4c4c0, 0x21)
    /root/.gvm/gos/go1.6.3/src/runtime/panic.go:547 +0x90 fp=0xca67b477f0     sp=0xca67b477d8
runtime.mapaccess1_faststr(0x2a8e520, 0xc9e29000f0, 0x2c11220, 0xa, 0x433e360)
    /root/.gvm/gos/go1.6.3/src/runtime/hashmap_fast.go:202 +0x5b fp=0xca67b47850 sp=0xca67b477f0
k8s.io/kubernetes/pkg/httplog.(*respLogger).Log(0xcbddf2ae70)
       /data/gerrit/src/k8s.io/kubernetes/_output/local/go/src/k8s.io/kubernetes/pkg/httplog/log.go:180 +0x43d fp=0xca67b47af8 sp=0xca67b47850
k8s.io/kubernetes/pkg/apiserver.RecoverPanics.func1(0x7f099f157090, 0xcbddf2ae70, 0xcd7569e380)
     /data/gerrit/src/k8s.io/kubernetes/_output/local/go/src/k8s.io/kubernetes/pkg/apiserver/handlers.go:174 +0x15d fp=0xca67b47b50 sp=0xca67b47af8
    net/http.HandlerFunc.ServeHTTP(0xc821a4eac0, 0x7f099f157058, 0xca0f4eb450, 0xcd7569e380)
    /root/.gvm/gos/go1.6.3/src/net/http/server.go:1618 +0x3a fp=0xca67b47b70 sp=0xca67b47b50
net/http.serverHandler.ServeHTTP(0xc8215a7b80, 0x7f099f157058, 0xca0f4eb450, 0xcd7569e380)
    /root/.gvm/gos/go1.6.3/src/net/http/server.go:2081 +0x19e fp=0xca67b47bd0 sp=0xca67b47b70
net/http.(*conn).serve(0xc8b5d6b980)
    /root/.gvm/gos/go1.6.3/src/net/http/server.go:1472 +0xf2e fp=0xca67b47f98 sp=0xca67b47bd0
runtime.goexit()
    /root/.gvm/gos/go1.6.3/src/runtime/asm_amd64.s:1998 +0x1 fp=0xca67b47fa0 sp=0xca67b47f98
created by net/http.(*Server).Serve
    /root/.gvm/gos/go1.6.3/src/net/http/server.go:2137 +0x44e

对应源码:

pkg/apiserver/handlers.go

145 func RecoverPanics(handler http.Handler) http.Handler {
146 return http.HandlerFunc(func(w http.ResponseWriter, req *http.Request) {
147 defer func() {
148     if x := recover(); x != nil {
149         http.Error(w, "apis panic. Look in log for details.", http.StatusInternalServerError)
150         glog.Errorf("APIServer panic'd on %v %v: %v\n%s\n", req.Method, req.RequestURI, x, debug.Stack())
151     }
152 }()
153 defer httplog.NewLogged(req, &w).StacktraceWhen(
        httplog.StatusIsNot(
            http.StatusOK,
            http.StatusCreated,
            http.StatusAccepted,
            http.StatusBadRequest,
            http.StatusMovedPermanently,
            http.StatusTemporaryRedirect,
            http.StatusConflict,
            http.StatusNotFound,
            http.StatusUnauthorized,
            http.StatusForbidden,
            errors.StatusUnprocessableEntity,
            http.StatusSwitchingProtocols,
            http.StatusRequestTimeout,
            errors.StatusTooManyRequests,
        ),
170 ).Log()

    // Dispatch to the internal handler
    handler.ServeHTTP(w, req)
174 })
}

pkg/httplog/log.go:

159 func (rl *respLogger) Log() {
160 latency := time.Since(rl.startTime)
161 if glog.V(2) {
162     extraInfo := ""
163     if latency >= time.Millisecond*200 && latency < time.Second {
        extraInfo = fmt.Sprintf("%d00.Millisecond", latency/(time.Millisecond*100))
    } else if latency >= time.Second && latency < time.Minute { // Warning
        extraInfo = fmt.Sprintf("%d.Second", latency/(time.Second))
    } else if latency >= time.Minute { // nce will timeout
        extraInfo = fmt.Sprintf("%d.Minutes", latency/(time.Minute))
    }
    method := rl.req.Method
    if len(rl.req.Header["Detailed-Method"]) > 0 {
        method = rl.req.Header["Detailed-Method"][0]
    }

    remoteIP := rl.getXForwardIPAdress(rl.req)


    if !rl.hijacked {
        //glog.InfoDepth(1, fmt.Sprintf("%s %s: (%v) %v%v%v [%s %s]", rl.req.Method, rl.req.RequestURI, latency, rl.status, rl.statusStack, rl.addedInfo, rl.req.Header["User-Agent"], rl.req.RemoteAddr))
180         glog.InfoDepth(1, fmt.Sprintf("%v %s %s: (%sms) %v%v [%s %s]-%s %v", rl.req.Header["X-Requestid"], method, rl.req.RequestURI, GetMilliLatency(latency), rl.status, rl.addedInfo, rl.req.Header["User-Agent"], remoteIP, extraInfo, rl.statusStack))
    } else {
        //glog.InfoDepth(1, fmt.Sprintf("%s %s: (%v) hijacked [%s %s]", rl.req.Method, rl.req.RequestURI, latency, rl.req.Header["User-Agent"], rl.req.RemoteAddr))
        glog.InfoDepth(1, fmt.Sprintf("%v %s %s: (%sms) hijacked [%s %s]-%s", rl.req.Header["X-Requestid"], method, rl.req.RequestURI, GetMilliLatency(latency), rl.req.Header["User-Agent"], remoteIP, extraInfo))
    }
}

在 handler.ServeHTTP 中，我更改了 request.Header。我找不到原因:

(1) “serveHTTP”和“defer log”以串行方式工作。所以不存在“并发读写”

(2) 即使存在“并发读写”，也存在“recover”函数来处理panic，所以kube-apiserver不应该退出。

这个问题耗费了我很多时间，谁能帮帮我？谢谢

Answer 1

我的猜测是您已经踩到了全局超时处理程序实际上无法停止处理请求的 goroutines 的事实。因此，当连接超时(1m，除非您已更改它)时，它会返回错误。如果 goroutine 实际上同时完成，它也可以尝试写入连接。我以为我们已经添加了一个锁来防止这引起问题，但也许 header 并不总是受到锁的保护。如果您可以在干净的源代码树中重现它，请在 Kubernetes github 存储库中提交问题。