go http 框架性能大幅下降原因分析

    最近在开发一个web 框架，然后业务方使用过程中，跟我们说，压测qps 上不去，我就很纳闷，httprouter + net/http.httpserver , 性能不可能这么差啊，网上的压测结果都是10w qps 以上，几个middleware 至于将性能拖垮？后来一番排查，发现些有意思的东西。

    首先，我就简单压测hello world, 每个请求进来，我日志都不打，然后，打开pprof ,显示的情况如下：

    这里futex 怎么这么高？看着上面的一些操作，addtimer, deltimer 我想到以前的自己实现的定时器，这估计是超时引起的。然后检查版本，go1.9,  然后框架默认为每个conn 设置了4个timeout，readtimeout, writetimeout, idletimeout, headertimeout ，这直接导致了定时器在添加和删除回调的时候，锁的压力特别大。

    下面我们分析下，正常的加超时操作，到底发生了些什么，下面是个最简单的例子，为了安全，每个连接设置超时。

package main

import (
"fmt"
"github.com/julienschmidt/httprouter"
"log"
"net/http"
"time"
)

func Index(w http.ResponseWriter, r *http.Request, _ httprouter.Params) {
fmt.Fprint(w, "Welcome!\n")
}

func Hello(w http.ResponseWriter, r *http.Request, ps httprouter.Params) {
fmt.Fprintf(w, "hello, %s!\n", ps.ByName("name"))
}

func main() {

router := httprouter.New()
router.GET("/", Index)
router.GET("/hello/:name", Hello)

srv := &http.Server{
ReadTimeout:       5 * time.Second,
WriteTimeout:      10 * time.Second,
ReadHeaderTimeout: 10 * time.Second,
IdleTimeout:       10 * time.Second,
Addr:              "0.0.0.0:8998",
Handler:           router,
}

log.Fatal(srv.ListenAndServe())
}

    其中，ListenAndServe() 在调用accept 每个连接后，会调用 server.serve(), 根据是否添加超时，调用conn.SetReadDeadline等函数，对应的是 net/http/server.go,如下：

// Serve a new connection.
func (c *conn) serve(ctx context.Context) {
...

if tlsConn, ok := c.rwc.(*tls.Conn); ok {
if d := c.server.ReadTimeout; d != 0 {
c.rwc.SetReadDeadline(time.Now().Add(d)) // 设置读超时
}
if d := c.server.WriteTimeout; d != 0 {
c.rwc.SetWriteDeadline(time.Now().Add(d))// 设置写超时
}
if err := tlsConn.Handshake(); err != nil {
c.server.logf("http: TLS handshake error from %s: %v", c.rwc.RemoteAddr(), err)
return
}
c.tlsState = new(tls.ConnectionState)
*c.tlsState = tlsConn.ConnectionState()
if proto := c.tlsState.NegotiatedProtocol; validNPN(proto) {
if fn := c.server.TLSNextProto[proto]; fn != nil {
h := initNPNRequest{tlsConn, serverHandler{c.server}}
fn(c.server, tlsConn, h)
}
return
}
}
...

    之后，con.SetReadDeadline 会调用 internal/poll/fd_poll_runtime.go的 fd.setReadDeadline，最后调用runtime/netpoll.go 的poll_runtime_pollSetDeadline， 这个函数会链接成internal/poll.runtime_pollSetDeadline。这个函数比较关键：

//go:linkname poll_runtime_pollSetDeadline internal/poll.runtime_pollSetDeadline
func poll_runtime_pollSetDeadline(pd *pollDesc, d int64, mode int) {
lock(&pd.lock)
if pd.closing {
unlock(&pd.lock)
return
}
pd.seq++ // invalidate current timers
// Reset current timers.
if pd.rt.f != nil {
deltimer(&pd.rt)
pd.rt.f = nil
}
if pd.wt.f != nil {
deltimer(&pd.wt)
pd.wt.f = nil
}
// Setup new timers.
if d != 0 && d <= nanotime() {
d = -1
}
if mode == 'r' || mode == 'r'+'w' {
pd.rd = d
}
if mode == 'w' || mode == 'r'+'w' {
pd.wd = d
}
if pd.rd > 0 && pd.rd == pd.wd {
pd.rt.f = netpollDeadline
pd.rt.when = pd.rd
// Copy current seq into the timer arg.
// Timer func will check the seq against current descriptor seq,
// if they differ the descriptor was reused or timers were reset.
pd.rt.arg = pd
pd.rt.seq = pd.seq
addtimer(&pd.rt)
} else {
if pd.rd > 0 {
pd.rt.f = netpollReadDeadline // 设置读的定时回调
pd.rt.when = pd.rd
pd.rt.arg = pd
pd.rt.seq = pd.seq
addtimer(&pd.rt)             // 添加到系统定时器中
}
if pd.wd > 0 {
pd.wt.f = netpollWriteDeadline // 设置写的定时回调
pd.wt.when = pd.wd
pd.wt.arg = pd
pd.wt.seq = pd.seq
addtimer(&pd.wt)             // 添加到系统定时器中
}
}
// If we set the new deadline in the past, unblock currently pending IO if any.
var rg, wg *g
atomicstorep(unsafe.Pointer(&wg), nil) // full memory barrier between stores to rd/wd and load of rg/wg in netpollunblock
if pd.rd < 0 {
rg = netpollunblock(pd, 'r', false)
}
if pd.wd < 0 {
wg = netpollunblock(pd, 'w', false)
}
unlock(&pd.lock)
if rg != nil {
netpollgoready(rg, 3)
}
if wg != nil {
netpollgoready(wg, 3)
}
}

    这里主要工作就是检查过期定时器，然后添加定时器，设置回调函数为netpollReadDeadline 或者netpollWriteDeadline。 从中可以看出添加和删除定时器操作为addtimer(&pd.rt), deltimer(&pd.rt)。

    后面就是核心了，为啥加超时后这么慢，看下addtimer 的实现，timer 是个四叉小顶堆，每次添加一个超时，最后都需要对一个全局的timers 进行加锁，当qps 很高，一个请求，多次加锁，这性能能很高吗？

type timer struct {
i int // heap index

// Timer wakes up at when, and then at when+period, ... (period > 0 only)
// each time calling f(arg, now) in the timer goroutine, so f must be
// a well-behaved function and not block.
when   int64
period int64
f      func(interface{}, uintptr)
arg    interface{}
seq    uintptr
}

var timers struct {
lock         mutex
gp           *g
created      bool
sleeping     bool
rescheduling bool
sleepUntil   int64
waitnote     note
t            []*timer
}

//添加一个定时器

func addtimer(t *timer) {
lock(&timers.lock)
addtimerLocked(t)
unlock(&timers.lock)
}

    解决锁冲突改怎么办？分段锁是很常见一个思路，在go1.10 后，timers 由一个，变成64个，定时器被打散到64个锁上去，自然锁冲突就降低了。看1.10的runtime/time.go 可以发现定义如下，每个p有单独的timer， 每个timer能被多个p使用：

// Package time knows the layout of this structure.
// If this struct changes, adjust ../time/sleep.go:/runtimeTimer.
// For GOOS=nacl, package syscall knows the layout of this structure.
// If this struct changes, adjust ../syscall/net_nacl.go:/runtimeTimer.
type timer struct {
tb *timersBucket // the bucket the timer lives in
i  int           // heap index

// Timer wakes up at when, and then at when+period, ... (period > 0 only)
// each time calling f(arg, now) in the timer goroutine, so f must be
// a well-behaved function and not block.
when   int64
period int64
f      func(interface{}, uintptr)
arg    interface{}
seq    uintptr
}

// timersLen is the length of timers array.
//
// Ideally, this would be set to GOMAXPROCS, but that would require
// dynamic reallocation
//
// The current value is a compromise between memory usage and performance
// that should cover the majority of GOMAXPROCS values used in the wild.
const timersLen = 64 //64个bucket

// timers contains "per-P" timer heaps.
//
// Timers are queued into timersBucket associated with the current P,
// so each P may work with its own timers independently of other P instances.
//
// Each timersBucket may be associated with multiple P
// if GOMAXPROCS > timersLen.
var timers [timersLen]struct {
timersBucket

// The padding should eliminate false sharing
// between timersBucket values.
pad [sys.CacheLineSize - unsafe.Sizeof(timersBucket{})%sys.CacheLineSize]byte
}

下面是go1.10 后的timer 数据结构（此图来源于网络）：

    总结，网上很多httpserver 框架压测 qps 很高，但是它们的demo并没有设置超时，数据真实值会差很多。线上如果需要设置超时，需要注意go 的版本，qps 很高的情况下，最好使用1.10以上。最终我们不做任何其他操作情况下，仅将go 版本提高到1.10，qps 提高接近2倍。