亲宝软件园·资讯

展开

一文带你了解Go语言实现的并发神库conc

asong2020 人气:0

前言

哈喽,大家好,我是asong;前几天逛github发现了一个有趣的并发库-conc,其目标是:

从简介上看主要封装功能如下:

接下来就区分模块来介绍一下这个库;

仓库地址:https://github.com/sourcegraph/conc

Go语言标准库有提供sync.waitGroup控制等待goroutine,我们一般会写出如下代码:

func main(){
    var wg sync.WaitGroup
    for i:=0; i < 10; i++{
        wg.Add(1)
        go func() {
            defer wg.Done()
            defer func() {
                // recover panic
                err := recover()
                if err != nil {
                    fmt.Println(err)
                }
            }
            // do something
            handle()
        }
    }
    wg.Wait()
}

上述代码我们需要些一堆重复代码,并且需要单独在每一个func中处理recover逻辑,所以conc库对其进行了封装,代码简化如下:

func main() {
	wg := conc.NewWaitGroup()
	for i := 0; i < 10; i++ {
		wg.Go(doSomething)
	}
	wg.Wait()
}

func doSomething() {
	fmt.Println("test")
}

conc库封装也比较简单,结构如下:

type WaitGroup struct {
	wg sync.WaitGroup
	pc panics.Catcher
}

其自己实现了Catcher类型对recover逻辑进行了封装,封装思路如下:

type Catcher struct {
	recovered atomic.Pointer[RecoveredPanic]
}

recovered是原子指针类型,RecoveredPanic是捕获的recover封装,封装了堆栈等信息:

type RecoveredPanic struct {
	// The original value of the panic.
	Value any
	// The caller list as returned by runtime.Callers when the panic was
	// recovered. Can be used to produce a more detailed stack information with
	// runtime.CallersFrames.
	Callers []uintptr
	// The formatted stacktrace from the goroutine where the panic was recovered.
	// Easier to use than Callers.
	Stack []byte
}

提供了Try方法执行方法,只会记录第一个panic的goroutine信息:

func (p *Catcher) Try(f func()) {
	defer p.tryRecover()
	f()
}

func (p *Catcher) tryRecover() {
	if val := recover(); val != nil {
		rp := NewRecoveredPanic(1, val)
        // 只会记录第一个panic的goroutine信息
		p.recovered.CompareAndSwap(nil, &rp)
	}
}

提供了Repanic()方法用来重放捕获的panic:

func (p *Catcher) Repanic() {
	if val := p.Recovered(); val != nil {
		panic(val)
	}
}

func (p *Catcher) Recovered() *RecoveredPanic {
	return p.recovered.Load()
}

waitGroup对此也分别提供了Wait()WaitAndRecover()方法:

func (h *WaitGroup) Wait() {
	h.wg.Wait()

	// Propagate a panic if we caught one from a child goroutine.
	h.pc.Repanic()
}

func (h *WaitGroup) WaitAndRecover() *panics.RecoveredPanic {
	h.wg.Wait()

	// Return a recovered panic if we caught one from a child goroutine.
	return h.pc.Recovered()
}

wait方法只要有一个goroutine发生panic就会向上抛出panic,比较简单粗暴;

waitAndRecover方法只有有一个goroutine发生panic就会返回第一个recover的goroutine信息;

总结:conc库对waitGrouop的封装总体是比较不错的,可以减少重复的代码;

worker池

conc提供了几种类型的worker池:

我们来看一个简单的例子:

import "github.com/sourcegraph/conc/pool"

func ExampleContextPool_WithCancelOnError() {
	p := pool.New().
		WithMaxGoroutines(4).
		WithContext(context.Background()).
		WithCancelOnError()
	for i := 0; i < 3; i++ {
		i := i
		p.Go(func(ctx context.Context) error {
			if i == 2 {
				return errors.New("I will cancel all other tasks!")
			}
			<-ctx.Done()
			return nil
		})
	}
	err := p.Wait()
	fmt.Println(err)
	// Output:
	// I will cancel all other tasks!
}

在创建pool时有如下方法可以调用:

pool的基础结构如下:

type Pool struct {
	handle   conc.WaitGroup
	limiter  limiter
	tasks    chan func()
	initOnce sync.Once
}

limiter是控制器,用chan来控制goroutine的数量:

type limiter chan struct{}

func (l limiter) limit() int {
	return cap(l)
}

func (l limiter) release() {
	if l != nil {
		<-l
	}
}

pool的核心逻辑也比较简单,如果没有设置limiter,那么就看有没有空闲的worker,否则就创建一个新的worker,然后投递任务进去;

如果设置了limiter,达到了limiter worker数量上限,就把任务投递给空闲的worker,没有空闲就阻塞等着;

func (p *Pool) Go(f func()) {
	p.init()

	if p.limiter == nil {
		// 没有限制
		select {
		case p.tasks <- f:
			// A goroutine was available to handle the task.
		default:
			// No goroutine was available to handle the task.
			// Spawn a new one and send it the task.
			p.handle.Go(p.worker)
			p.tasks <- f
		}
	} else {
		select {
		case p.limiter <- struct{}{}:
			// If we are below our limit, spawn a new worker rather
			// than waiting for one to become available.
			p.handle.Go(p.worker)

			// We know there is at least one worker running, so wait
			// for it to become available. This ensures we never spawn
			// more workers than the number of tasks.
			p.tasks <- f
		case p.tasks <- f:
			// A worker is available and has accepted the task.
			return
		}
	}

}

这里work使用的是一个无缓冲的channel,这种复用方式很巧妙,如果goroutine执行很快避免创建过多的goroutine;

使用pool处理任务不能保证有序性,conc库又提供了Stream方法,返回结果可以保持顺序;

Stream

Steam的实现也是依赖于pool,在此基础上做了封装保证结果的顺序性,先看一个例子:

func ExampleStream() {
	times := []int{20, 52, 16, 45, 4, 80}

	stream := stream2.New()
	for _, millis := range times {
		dur := time.Duration(millis) * time.Millisecond
		stream.Go(func() stream2.Callback {
			time.Sleep(dur)
			// This will print in the order the tasks were submitted
			return func() { fmt.Println(dur) }
		})
	}
	stream.Wait()

	// Output:
	// 20ms
	// 52ms
	// 16ms
	// 45ms
	// 4ms
	// 80ms
}

stream的结构如下:

type Stream struct {
	pool             pool.Pool
	callbackerHandle conc.WaitGroup
	queue            chan callbackCh

	initOnce sync.Once
}

queue是一个channel类型,callbackCh也是channel类型 - chan func():

type callbackCh chan func()

在提交goroutine时按照顺序生成callbackCh传递结果:

func (s *Stream) Go(f Task) {
	s.init()

	// Get a channel from the cache.
	ch := getCh()

	// Queue the channel for the callbacker.
	s.queue <- ch

	// Submit the task for execution.
	s.pool.Go(func() {
		defer func() {
			// In the case of a panic from f, we don't want the callbacker to
			// starve waiting for a callback from this channel, so give it an
			// empty callback.
			if r := recover(); r != nil {
				ch <- func() {}
				panic(r)
			}
		}()

		// Run the task, sending its callback down this task's channel.
		callback := f()
		ch <- callback
	})
}

var callbackChPool = sync.Pool{
	New: func() any {
		return make(callbackCh, 1)
	},
}

func getCh() callbackCh {
	return callbackChPool.Get().(callbackCh)
}

func putCh(ch callbackCh) {
	callbackChPool.Put(ch)
}

ForEach和map

ForEach

conc库提供了ForEach方法可以优雅的并发处理切片,看一下官方的例子:

conc库使用泛型进行了封装,我们只需要关注handle代码即可,避免冗余代码,我们自己动手写一个例子:

func main() {
	input := []int{1, 2, 3, 4}
	iterator := iter.Iterator[int]{
		MaxGoroutines: len(input) / 2,
	}

	iterator.ForEach(input, func(v *int) {
		if *v%2 != 0 {
			*v = -1
		}
	})

	fmt.Println(input)
}

ForEach内部实现为Iterator结构及核心逻辑如下:

type Iterator[T any] struct {
	MaxGoroutines int
}
func (iter Iterator[T]) ForEachIdx(input []T, f func(int, *T)) {
	if iter.MaxGoroutines == 0 {
		// iter is a value receiver and is hence safe to mutate
		iter.MaxGoroutines = defaultMaxGoroutines()
	}

	numInput := len(input)
	if iter.MaxGoroutines > numInput {
		// No more concurrent tasks than the number of input items.
		iter.MaxGoroutines = numInput
	}

	var idx atomic.Int64
	// 通过atomic控制仅创建一个闭包
	task := func() {
		i := int(idx.Add(1) - 1)
		for ; i < numInput; i = int(idx.Add(1) - 1) {
			f(i, &input[i])
		}
	}

	var wg conc.WaitGroup
	for i := 0; i < iter.MaxGoroutines; i++ {
		wg.Go(task)
	}
	wg.Wait()
}

可以设置并发的goroutine数量,默认取的是GOMAXPROCS ,也可以自定义传参;

并发执行这块设计的很巧妙,仅创建了一个闭包,通过atomic控制idx,避免频繁触发GC;

map

conc库提供的map方法可以得到对切片中元素结果,官方例子:

使用map可以提高代码的可读性,并且减少了冗余代码,自己写个例子:

func main() {
	input := []int{1, 2, 3, 4}
	mapper := iter.Mapper[int, bool]{
		MaxGoroutines: len(input) / 2,
	}

	results := mapper.Map(input, func(v *int) bool { return *v%2 == 0 })
	fmt.Println(results)
	// Output:
	// [false true false true]
}

map的实现也依赖于Iterator,也是调用的ForEachIdx方法,区别于ForEach是记录处理结果;

总结

花了小半天时间看了一下这个库,很多设计点值得我们学习,总结一下我学习到的知识点:

加载全部内容

相关教程
猜你喜欢
用户评论