以太坊原始碼分析(49）p2p-table.go原始碼分析

尹成發表於2018-05-14
原始碼Go
table.go主要實現了p2p的Kademlia協議。

### Kademlia協議簡介(建議閱讀references裡面的pdf文件)
Kademlia協議（以下簡稱Kad） 是美國紐約大學的PetarP. Maymounkov和David Mazieres.
在2002年釋出的一項研究結果《Kademlia: A peerto -peer information system based on
the XOR metric》。
簡單的說， Kad 是一種分散式雜湊表（ DHT） 技術， 不過和其他 DHT 實現技術比較，如
Chord、 CAN、 Pastry 等， Kad 通過獨特的以異或演算法（ XOR）為距離度量基礎，建立了一種
全新的 DHT 拓撲結構，相比於其他演算法，大大提高了路由查詢速度。


### table的結構和欄位

    const (
        alpha = 3 // Kademlia concurrency factor
        bucketSize = 16 // Kademlia bucket size
        hashBits = len(common.Hash{}) * 8
        nBuckets = hashBits + 1 // Number of buckets
    
        maxBondingPingPongs = 16
        maxFindnodeFailures = 5
    
        autoRefreshInterval = 1 * time.Hour
        seedCount = 30
        seedMaxAge = 5 * 24 * time.Hour
    )
    
    type Table struct {
        mutex sync.Mutex // protects buckets, their content, and nursery
        buckets [nBuckets]*bucket // index of known nodes by distance
        nursery []*Node // bootstrap nodes
        db *nodeDB // database of known nodes
    
        refreshReq chan chan struct{}
        closeReq chan struct{}
        closed chan struct{}
    
        bondmu sync.Mutex
        bonding map[NodeID]*bondproc
        bondslots chan struct{} // limits total number of active bonding processes
    
        nodeAddedHook func(*Node) // for testing
    
        net transport
        self *Node // metadata of the local node
    }


### 初始化


    func newTable(t transport, ourID NodeID, ourAddr *net.UDPAddr, nodeDBPath string) (*Table, error) {
        // If no node database was given, use an in-memory one
        //這個在之前的database.go裡面有介紹。 開啟leveldb。如果path為空。那麼開啟一個基於記憶體的db
        db, err := newNodeDB(nodeDBPath, Version, ourID)
        if err != nil {
            return nil, err
        }
        tab := &Table{
            net: t,
            db: db,
            self: NewNode(ourID, ourAddr.IP, uint16(ourAddr.Port), uint16(ourAddr.Port)),
            bonding: make(map[NodeID]*bondproc),
            bondslots: make(chan struct{}, maxBondingPingPongs),
            refreshReq: make(chan chan struct{}),
            closeReq: make(chan struct{}),
            closed: make(chan struct{}),
        }
        for i := 0; i < cap(tab.bondslots); i++ {
            tab.bondslots <- struct{}{}
        }
        for i := range tab.buckets {
            tab.buckets[i] = new(bucket)
        }
        go tab.refreshLoop()
        return tab, nil
    }

上面的初始化啟動了一個goroutine refreshLoop()，這個函式主要完成以下的工作。

1. 每一個小時進行一次重新整理工作(autoRefreshInterval)
2. 如果接收到refreshReq請求。那麼進行重新整理工作。
3. 如果接收到關閉訊息。那麼進行關閉。

所以函式主要的工作就是啟動重新整理工作。doRefresh


    // refreshLoop schedules doRefresh runs and coordinates shutdown.
    func (tab *Table) refreshLoop() {
        var (
            timer = time.NewTicker(autoRefreshInterval)
            waiting []chan struct{} // accumulates waiting callers while doRefresh runs
            done chan struct{} // where doRefresh reports completion
        )
    loop:
        for {
            select {
            case <-timer.C:
                if done == nil {
                    done = make(chan struct{})
                    go tab.doRefresh(done)
                }
            case req := <-tab.refreshReq:
                waiting = append(waiting, req)
                if done == nil {
                    done = make(chan struct{})
                    go tab.doRefresh(done)
                }
            case <-done:
                for _, ch := range waiting {
                    close(ch)
                }
                waiting = nil
                done = nil
            case <-tab.closeReq:
                break loop
            }
        }
    
        if tab.net != nil {
            tab.net.close()
        }
        if done != nil {
            <-done
        }
        for _, ch := range waiting {
            close(ch)
        }
        tab.db.close()
        close(tab.closed)
    }


doRefresh函式

    // doRefresh performs a lookup for a random target to keep buckets
    // full. seed nodes are inserted if the table is empty (initial
    // bootstrap or discarded faulty peers).
    // doRefresh 隨機查詢一個目標，以便保持buckets是滿的。如果table是空的，那麼種子節點會插入。 （比如最開始的啟動或者是刪除錯誤的節點之後）
    func (tab *Table) doRefresh(done chan struct{}) {
        defer close(done)
    
        // The Kademlia paper specifies that the bucket refresh should
        // perform a lookup in the least recently used bucket. We cannot
        // adhere to this because the findnode target is a 512bit value
        // (not hash-sized) and it is not easily possible to generate a
        // sha3 preimage that falls into a chosen bucket.
        // We perform a lookup with a random target instead.
        //這裡暫時沒看懂
        var target NodeID
        rand.Read(target[:])
        result := tab.lookup(target, false) //lookup是查詢距離target最近的k個節點
        if len(result) > 0 { //如果結果不為0 說明表不是空的，那麼直接返回。
            return
        }
    
        // The table is empty. Load nodes from the database and insert
        // them. This should yield a few previously seen nodes that are
        // (hopefully) still alive.
        //querySeeds函式在database.go章節有介紹，從資料庫裡面隨機的查詢可用的種子節點。
        //在最開始啟動的時候資料庫是空白的。也就是最開始的時候這個seeds返回的是空的。
        seeds := tab.db.querySeeds(seedCount, seedMaxAge)
        //呼叫bondall函式。會嘗試聯絡這些節點，並插入到表中。
        //tab.nursery是在命令列中指定的種子節點。
        //最開始啟動的時候。 tab.nursery的值是內建在程式碼裡面的。 這裡是有值的。
        //C:\GOPATH\src\github.com\ethereum\go-ethereum\mobile\params.go
        //這裡面寫死了值。 這個值是通過SetFallbackNodes方法寫入的。 這個方法後續會分析。
        //這裡會進行雙向的pingpong交流。 然後把結果儲存在資料庫。
        seeds = tab.bondall(append(seeds, tab.nursery...))
    
        if len(seeds) == 0 { //沒有種子節點被發現， 可能需要等待下一次重新整理。
            log.Debug("No discv4 seed nodes found")
        }
        for _, n := range seeds {
            age := log.Lazy{Fn: func() time.Duration { return time.Since(tab.db.lastPong(n.ID)) }}
            log.Trace("Found seed node in database", "id", n.ID, "addr", n.addr(), "age", age)
        }
        tab.mutex.Lock()
        //這個方法把所有經過bond的seed加入到bucket(前提是bucket未滿)
        tab.stuff(seeds)
        tab.mutex.Unlock()
    
        // Finally, do a self lookup to fill up the buckets.
        tab.lookup(tab.self.ID, false) // 有了種子節點。那麼查詢自己來填充buckets。
    }

bondall方法，這個方法就是多執行緒的呼叫bond方法。

    // bondall bonds with all given nodes concurrently and returns
    // those nodes for which bonding has probably succeeded.
    func (tab *Table) bondall(nodes []*Node) (result []*Node) {
        rc := make(chan *Node, len(nodes))
        for i := range nodes {
            go func(n *Node) {
                nn, _ := tab.bond(false, n.ID, n.addr(), uint16(n.TCP))
                rc <- nn
            }(nodes[i])
        }
        for range nodes {
            if n := <-rc; n != nil {
                result = append(result, n)
            }
        }
        return result
    }

bond方法。記得在udp.go中。當我們收到一個ping方法的時候，也有可能會呼叫這個方法


    // bond ensures the local node has a bond with the given remote node.
    // It also attempts to insert the node into the table if bonding succeeds.
    // The caller must not hold tab.mutex.
    // bond確保本地節點與給定的遠端節點具有繫結。(遠端的ID和遠端的IP)。
    // 如果繫結成功，它也會嘗試將節點插入表中。呼叫者必須持有tab.mutex鎖
    // A bond is must be established before sending findnode requests.
    // Both sides must have completed a ping/pong exchange for a bond to
    // exist. The total number of active bonding processes is limited in
    // order to restrain network use.
    // 傳送findnode請求之前必須建立一個繫結。  雙方為了完成一個bond必須完成雙向的ping/pong過程。
    // 為了節約網路資源。 同時存在的bonding處理流程的總數量是受限的。  
    // bond is meant to operate idempotently in that bonding with a remote
    // node which still remembers a previously established bond will work.
    // The remote node will simply not send a ping back, causing waitping
    // to time out.
    // bond 是冪等的操作，跟一個任然記得之前的bond的遠端節點進行bond也可以完成。 遠端節點會簡單的不會傳送ping。 等待waitping超時。
    // If pinged is true, the remote node has just pinged us and one half
    // of the process can be skipped.
    //  如果pinged是true。 那麼遠端節點已經給我們傳送了ping訊息。這樣一半的流程可以跳過。
    func (tab *Table) bond(pinged bool, id NodeID, addr *net.UDPAddr, tcpPort uint16) (*Node, error) {
        if id == tab.self.ID {
            return nil, errors.New("is self")
        }
        // Retrieve a previously known node and any recent findnode failures
        node, fails := tab.db.node(id), 0
        if node != nil {
            fails = tab.db.findFails(id)
        }
        // If the node is unknown (non-bonded) or failed (remotely unknown), bond from scratch
        var result error
        age := time.Since(tab.db.lastPong(id))
        if node == nil || fails > 0 || age > nodeDBNodeExpiration {
            //如果資料庫沒有這個節點。 或者錯誤數量大於0或者節點超時。
            log.Trace("Starting bonding ping/pong", "id", id, "known", node != nil, "failcount", fails, "age", age)
    
            tab.bondmu.Lock()
            w := tab.bonding[id]
            if w != nil {
                // Wait for an existing bonding process to complete.
                tab.bondmu.Unlock()
                <-w.done
            } else {
                // Register a new bonding process.
                w = &bondproc{done: make(chan struct{})}
                tab.bonding[id] = w
                tab.bondmu.Unlock()
                // Do the ping/pong. The result goes into w.
                tab.pingpong(w, pinged, id, addr, tcpPort)
                // Unregister the process after it's done.
                tab.bondmu.Lock()
                delete(tab.bonding, id)
                tab.bondmu.Unlock()
            }
            // Retrieve the bonding results
            result = w.err
            if result == nil {
                node = w.n
            }
        }
        if node != nil {
            // Add the node to the table even if the bonding ping/pong
            // fails. It will be relaced quickly if it continues to be
            // unresponsive.
            //這個方法比較重要。 如果對應的bucket有空間，會直接插入buckets。如果buckets滿了。 會用ping操作來測試buckets中的節點試圖騰出空間。
            tab.add(node)
            tab.db.updateFindFails(id, 0)
        }
        return node, result
    }

pingpong方法

    func (tab *Table) pingpong(w *bondproc, pinged bool, id NodeID, addr *net.UDPAddr, tcpPort uint16) {
        // Request a bonding slot to limit network usage
        <-tab.bondslots
        defer func() { tab.bondslots <- struct{}{} }()
    
        // Ping the remote side and wait for a pong.
        // Ping遠端節點。並等待一個pong訊息
        if w.err = tab.ping(id, addr); w.err != nil {
            close(w.done)
            return
        }
        //這個在udp收到一個ping訊息的時候被設定為真。這個時候我們已經收到對方的ping訊息了。
        //那麼我們就不同等待ping訊息了。 否則需要等待對方傳送過來的ping訊息(我們主動發起ping訊息)。
        if !pinged {
            // Give the remote node a chance to ping us before we start
            // sending findnode requests. If they still remember us,
            // waitping will simply time out.
            tab.net.waitping(id)
        }
        // Bonding succeeded, update the node database.
        // 完成bond過程。 把節點插入資料庫。 資料庫操作在這裡完成。 bucket的操作在tab.add裡面完成。 buckets是記憶體的操作。 資料庫是持久化的seeds節點。用來加速啟動過程的。
        w.n = NewNode(id, addr.IP, uint16(addr.Port), tcpPort)
        tab.db.updateNode(w.n)
        close(w.done)
    }

tab.add方法

    // add attempts to add the given node its corresponding bucket. If the
    // bucket has space available, adding the node succeeds immediately.
    // Otherwise, the node is added if the least recently active node in
    // the bucket does not respond to a ping packet.
    // add試圖把給定的節點插入對應的bucket。 如果bucket有空間，那麼直接插入。 否則，如果bucket中最近活動的節點沒有響應ping操作，那麼我們就使用這個節點替換它。
    // The caller must not hold tab.mutex.
    func (tab *Table) add(new *Node) {
        b := tab.buckets[logdist(tab.self.sha, new.sha)]
        tab.mutex.Lock()
        defer tab.mutex.Unlock()
        if b.bump(new) { //如果節點存在。那麼更新它的值。然後退出。
            return
        }
        var oldest *Node
        if len(b.entries) == bucketSize {
            oldest = b.entries[bucketSize-1]
            if oldest.contested {
                // The node is already being replaced, don't attempt
                // to replace it.
                // 如果別的goroutine正在對這個節點進行測試。 那麼取消替換， 直接退出。
                // 因為ping的時間比較長。所以這段時間是沒有加鎖的。 用了contested這個狀態來標識這種情況。
                return
            }
            oldest.contested = true
            // Let go of the mutex so other goroutines can access
            // the table while we ping the least recently active node.
            tab.mutex.Unlock()
            err := tab.ping(oldest.ID, oldest.addr())
            tab.mutex.Lock()
            oldest.contested = false
            if err == nil {
                // The node responded, don't replace it.
                return
            }
        }
        added := b.replace(new, oldest)
        if added && tab.nodeAddedHook != nil {
            tab.nodeAddedHook(new)
        }
    }



stuff方法比較簡單。 找到對應節點應該插入的bucket。 如果這個bucket沒有滿，那麼就插入這個bucket。否則什麼也不做。 需要說一下的是logdist()這個方法。這個方法對兩個值進行按照位置異或，然後返回最高位的下標。 比如 logdist(101,010) = 3 logdist(100, 100) = 0 logdist(100,110) = 2

    // stuff adds nodes the table to the end of their corresponding bucket
    // if the bucket is not full. The caller must hold tab.mutex.
    func (tab *Table) stuff(nodes []*Node) {
    outer:
        for _, n := range nodes {
            if n.ID == tab.self.ID {
                continue // don't add self
            }
            bucket := tab.buckets[logdist(tab.self.sha, n.sha)]
            for i := range bucket.entries {
                if bucket.entries[i].ID == n.ID {
                    continue outer // already in bucket
                }
            }
            if len(bucket.entries) < bucketSize {
                bucket.entries = append(bucket.entries, n)
                if tab.nodeAddedHook != nil {
                    tab.nodeAddedHook(n)
                }
            }
        }
    }


在看看之前的Lookup函式。 這個函式用來查詢一個指定節點的資訊。 這個函式首先從本地拿到距離這個節點最近的所有16個節點。 然後給所有的節點傳送findnode的請求。 然後對返回的界定進行bondall處理。 然後返回所有的節點。



    func (tab *Table) lookup(targetID NodeID, refreshIfEmpty bool) []*Node {
        var (
            target = crypto.Keccak256Hash(targetID[:])
            asked = make(map[NodeID]bool)
            seen = make(map[NodeID]bool)
            reply = make(chan []*Node, alpha)
            pendingQueries = 0
            result *nodesByDistance
        )
        // don't query further if we hit ourself.
        // unlikely to happen often in practice.
        asked[tab.self.ID] = true
        不會詢問我們自己
        for {
            tab.mutex.Lock()
            // generate initial result set
            result = tab.closest(target, bucketSize)
            //求取和target最近的16個節點
            tab.mutex.Unlock()
            if len(result.entries) > 0 || !refreshIfEmpty {
                break
            }
            // The result set is empty, all nodes were dropped, refresh.
            // We actually wait for the refresh to complete here. The very
            // first query will hit this case and run the bootstrapping
            // logic.
            <-tab.refresh()
            refreshIfEmpty = false
        }
    
        for {
            // ask the alpha closest nodes that we haven't asked yet
            // 這裡會併發的查詢，每次3個goroutine併發(通過pendingQueries引數進行控制)
            // 每次迭代會查詢result中和target距離最近的三個節點。
            for i := 0; i < len(result.entries) && pendingQueries < alpha; i++ {
                n := result.entries[i]
                if !asked[n.ID] { //如果沒有查詢過 //因為這個result.entries會被重複迴圈很多次。 所以用這個變數控制那些已經處理過了。
                    asked[n.ID] = true
                    pendingQueries++
                    go func() {
                        // Find potential neighbors to bond with
                        r, err := tab.net.findnode(n.ID, n.addr(), targetID)
                        if err != nil {
                            // Bump the failure counter to detect and evacuate non-bonded entries
                            fails := tab.db.findFails(n.ID) + 1
                            tab.db.updateFindFails(n.ID, fails)
                            log.Trace("Bumping findnode failure counter", "id", n.ID, "failcount", fails)
    
                            if fails >= maxFindnodeFailures {
                                log.Trace("Too many findnode failures, dropping", "id", n.ID, "failcount", fails)
                                tab.delete(n)
                            }
                        }
                        reply <- tab.bondall(r)
                    }()
                }
            }
            if pendingQueries == 0 {
                // we have asked all closest nodes, stop the search
                break
            }
            // wait for the next reply
            for _, n := range <-reply {
                if n != nil && !seen[n.ID] { //因為不同的遠方節點可能返回相同的節點。所有用seen[]來做排重。
                    seen[n.ID] = true
                    //這個地方需要注意的是, 查詢出來的結果又會加入result這個佇列。也就是說這是一個迴圈查詢的過程， 只要result裡面不斷加入新的節點。這個迴圈就不會終止。
                    result.push(n, bucketSize)
                }
            }
            pendingQueries--
        }
        return result.entries
    }
    
    // closest returns the n nodes in the table that are closest to the
    // given id. The caller must hold tab.mutex.
    func (tab *Table) closest(target common.Hash, nresults int) *nodesByDistance {
        // This is a very wasteful way to find the closest nodes but
        // obviously correct. I believe that tree-based buckets would make
        // this easier to implement efficiently.
        close := &nodesByDistance{target: target}
        for _, b := range tab.buckets {
            for _, n := range b.entries {
                close.push(n, nresults)
            }
        }
        return close
    }

result.push方法，這個方法會根據 所有的節點對於target的距離進行排序。 按照從近到遠的方式決定新節點的插入順序。(佇列中最大會包含16個元素)。 這樣會導致佇列裡面的元素和target的距離越來越近。距離相對遠的會被踢出佇列。
    
    // nodesByDistance is a list of nodes, ordered by
    // distance to target.
    type nodesByDistance struct {
        entries []*Node
        target common.Hash
    }
    
    // push adds the given node to the list, keeping the total size below maxElems.
    func (h *nodesByDistance) push(n *Node, maxElems int) {
        ix := sort.Search(len(h.entries), func(i int) bool {
            return distcmp(h.target, h.entries[i].sha, n.sha) > 0
        })
        if len(h.entries) < maxElems {
            h.entries = append(h.entries, n)
        }
        if ix == len(h.entries) {
            // farther away than all nodes we already have.
            // if there was room for it, the node is now the last element.
        } else {
            // slide existing entries down to make room
            // this will overwrite the entry we just appended.
            copy(h.entries[ix+1:], h.entries[ix:])
            h.entries[ix] = n
        }
    }


### table.go 匯出的一些方法
Resolve方法和Lookup方法

    // Resolve searches for a specific node with the given ID.
    // It returns nil if the node could not be found.
    //Resolve方法用來獲取一個指定ID的節點。 如果節點在本地。那麼返回本地節點。 否則執行
    //Lookup在網路上查詢一次。 如果查詢到節點。那麼返回。否則返回nil
    func (tab *Table) Resolve(targetID NodeID) *Node {
        // If the node is present in the local table, no
        // network interaction is required.
        hash := crypto.Keccak256Hash(targetID[:])
        tab.mutex.Lock()
        cl := tab.closest(hash, 1)
        tab.mutex.Unlock()
        if len(cl.entries) > 0 && cl.entries[0].ID == targetID {
            return cl.entries[0]
        }
        // Otherwise, do a network lookup.
        result := tab.Lookup(targetID)
        for _, n := range result {
            if n.ID == targetID {
                return n
            }
        }
        return nil
    }
    
    // Lookup performs a network search for nodes close
    // to the given target. It approaches the target by querying
    // nodes that are closer to it on each iteration.
    // The given target does not need to be an actual node
    // identifier.
    func (tab *Table) Lookup(targetID NodeID) []*Node {
        return tab.lookup(targetID, true)
    }

SetFallbackNodes方法，這個方法設定初始化的聯絡節點。 在table是空而且資料庫裡面也沒有已知的節點，這些節點可以幫助連線上網路，

    // SetFallbackNodes sets the initial points of contact. These nodes
    // are used to connect to the network if the table is empty and there
    // are no known nodes in the database.
    func (tab *Table) SetFallbackNodes(nodes []*Node) error {
        for _, n := range nodes {
            if err := n.validateComplete(); err != nil {
                return fmt.Errorf("bad bootstrap/fallback node %q (%v)", n, err)
            }
        }
        tab.mutex.Lock()
        tab.nursery = make([]*Node, 0, len(nodes))
        for _, n := range nodes {
            cpy := *n
            // Recompute cpy.sha because the node might not have been
            // created by NewNode or ParseNode.
            cpy.sha = crypto.Keccak256Hash(n.ID[:])
            tab.nursery = append(tab.nursery, &cpy)
        }
        tab.mutex.Unlock()
        tab.refresh()
        return nil
    }


### 總結

這樣， p2p網路的Kademlia協議就完結了。 基本上是按照論文進行實現。 udp進行網路通訊。資料庫儲存連結過的節點。 table實現了Kademlia的核心。 根據異或距離來進行節點的查詢。 節點的發現和更新等流程。




網址：http://www.qukuailianxueyuan.io/



欲領取造幣技術與全套虛擬機器資料

區塊鏈技術交流QQ群：756146052  備註：CSDN

尹成學院微信：備註：CSDN



相關文章