Redisson-關於使用訂閱數問題

玉樹臨楓發表於2022-01-13

一、前提

最近在使用分散式鎖redisson時遇到一個線上問題:發現是subscriptionsPerConnection or subscriptionConnectionPoolSize 的大小不夠,需要提高配置才能解決。

二、原始碼分析

下面對其原始碼進行分析,才能找到到底是什麼邏輯導致問題所在:

1、RedissonLock#lock() 方法

private void lock(long leaseTime, TimeUnit unit, boolean interruptibly) throws InterruptedException {
        long threadId = Thread.currentThread().getId();
        // 嘗試獲取,如果ttl == null,則表示獲取鎖成功
        Long ttl = tryAcquire(leaseTime, unit, threadId);
        // lock acquired
        if (ttl == null) {
            return;
        }

        // 訂閱鎖釋放事件,並通過await方法阻塞等待鎖釋放,有效的解決了無效的鎖申請浪費資源的問題
        RFuture<RedissonLockEntry> future = subscribe(threadId);
        if (interruptibly) {
            commandExecutor.syncSubscriptionInterrupted(future);
        } else {
            commandExecutor.syncSubscription(future);
        }

        // 後面程式碼忽略
        try {
            // 無限迴圈獲取鎖,直到獲取鎖成功
            // ...
        } finally {
            // 取消訂閱鎖釋放事件
            unsubscribe(future, threadId);
        }
}

總結下主要邏輯:

  1. 獲取當前執行緒的執行緒id;
  2. tryAquire嘗試獲取鎖,並返回ttl
  3. 如果ttl為空,則結束流程;否則進入後續邏輯;
  4. this.subscribe(threadId)訂閱當前執行緒,返回一個RFuture;
  5. 如果在指定時間沒有監聽到,則會產生如上異常。
  6. 訂閱成功後, 通過while(true)迴圈,一直嘗試獲取鎖
  7. fially程式碼塊,會解除訂閱
    所以上述這情況問題應該出現在subscribe()方法中

2、詳細看下subscribe()方法

protected RFuture<RedissonLockEntry> subscribe(long threadId) {
    // entryName 格式:“id:name”;
    // channelName 格式:“redisson_lock__channel:name”;
    return pubSub.subscribe(getEntryName(), getChannelName());
}

RedissonLock#pubSub 是在RedissonLock建構函式中初始化的:

public RedissonLock(CommandAsyncExecutor commandExecutor, String name) {
    // ....
    this.pubSub = commandExecutor.getConnectionManager().getSubscribeService().getLockPubSub();
}

subscribeServiceMasterSlaveConnectionManager的實現中又是通過如下方式構造的

public MasterSlaveConnectionManager(MasterSlaveServersConfig cfg, Config config, UUID id) {
    this(config, id);
    this.config = cfg;

    // 初始化
    initTimer(cfg);
    initSingleEntry();
}

protected void initTimer(MasterSlaveServersConfig config) {
    int[] timeouts = new int[]{config.getRetryInterval(), config.getTimeout()};
    Arrays.sort(timeouts);
    int minTimeout = timeouts[0];
    if (minTimeout % 100 != 0) {
        minTimeout = (minTimeout % 100) / 2;
    } else if (minTimeout == 100) {
        minTimeout = 50;
    } else {
        minTimeout = 100;
    }

    timer = new HashedWheelTimer(new DefaultThreadFactory("redisson-timer"), minTimeout, TimeUnit.MILLISECONDS, 1024, false);

    connectionWatcher = new IdleConnectionWatcher(this, config);

    // 初始化:其中this就是MasterSlaveConnectionManager例項,config則為MasterSlaveServersConfig例項:
    subscribeService = new PublishSubscribeService(this, config);
}

PublishSubscribeService建構函式

private final SemaphorePubSub semaphorePubSub = new SemaphorePubSub(this);
public PublishSubscribeService(ConnectionManager connectionManager, MasterSlaveServersConfig config) {
    super();
    this.connectionManager = connectionManager;
    this.config = config;
    for (int i = 0; i < locks.length; i++) {
        // 這裡初始化了一組訊號量,每個訊號量的初始值為1
        locks[i] = new AsyncSemaphore(1);
    }
}

3、回到subscribe()方法主要邏輯還是交給了 LockPubSub#subscribe()裡面

private final ConcurrentMap<String, E> entries = new ConcurrentHashMap<>();

public RFuture<E> subscribe(String entryName, String channelName) {
      // 從PublishSubscribeService獲取對應的訊號量。 相同的channelName獲取的是同一個訊號量
     // public AsyncSemaphore getSemaphore(ChannelName channelName) {
    //    return locks[Math.abs(channelName.hashCode() % locks.length)];
    // }
    AsyncSemaphore semaphore = service.getSemaphore(new ChannelName(channelName));

    AtomicReference<Runnable> listenerHolder = new AtomicReference<Runnable>();    
    RPromise<E> newPromise = new RedissonPromise<E>() {
        @Override
        public boolean cancel(boolean mayInterruptIfRunning) {
            return semaphore.remove(listenerHolder.get());
        }
    };

    Runnable listener = new Runnable() {

        @Override
        public void run() {
            //  如果存在RedissonLockEntry, 則直接利用已有的監聽
            E entry = entries.get(entryName);
            if (entry != null) {
                entry.acquire();
                semaphore.release();
                entry.getPromise().onComplete(new TransferListener<E>(newPromise));
                return;
            }

            E value = createEntry(newPromise);
            value.acquire();

            E oldValue = entries.putIfAbsent(entryName, value);
            if (oldValue != null) {
                oldValue.acquire();
                semaphore.release();
                oldValue.getPromise().onComplete(new TransferListener<E>(newPromise));
                return;
            }

            // 建立監聽,
            RedisPubSubListener<Object> listener = createListener(channelName, value);
            // 訂閱監聽
            service.subscribe(LongCodec.INSTANCE, channelName, semaphore, listener);
        }
    };

    // 最終會執行listener.run方法
    semaphore.acquire(listener);
    listenerHolder.set(listener);

    return newPromise;
}

AsyncSemaphore#acquire()方法

public void acquire(Runnable listener) {
    acquire(listener, 1);
}

public void acquire(Runnable listener, int permits) {
    boolean run = false;

    synchronized (this) {
        // counter初始化值為1
        if (counter < permits) {
            // 如果不是第一次執行,則將listener加入到listeners集合中
            listeners.add(new Entry(listener, permits));
            return;
        } else {
            counter -= permits;
            run = true;
        }
    }

    // 第一次執行acquire, 才會執行listener.run()方法
    if (run) {
        listener.run();
    }
}

梳理上述邏輯:

  • 1、從PublishSubscribeService獲取對應的訊號量, 相同的channelName獲取的是同一個訊號量
  • 2、如果是第一次請求,則會立馬執行listener.run()方法, 否則需要等上個執行緒獲取到該訊號量執行完方能執行;
  • 3、如果已經存在RedissonLockEntry, 則利用已經訂閱就行
  • 4、如果不存在RedissonLockEntry, 則會建立新的RedissonLockEntry,然後進行。
    從上面程式碼看,主要邏輯是交給了PublishSubscribeService#subscribe方法

4、PublishSubscribeService#subscribe邏輯如下:

private final ConcurrentMap<ChannelName, PubSubConnectionEntry> name2PubSubConnection = new ConcurrentHashMap<>();
private final Queue<PubSubConnectionEntry> freePubSubConnections = new ConcurrentLinkedQueue<>();

public RFuture<PubSubConnectionEntry> subscribe(Codec codec, String channelName, AsyncSemaphore semaphore, RedisPubSubListener<?>... listeners) {
    RPromise<PubSubConnectionEntry> promise = new RedissonPromise<PubSubConnectionEntry>();
    // 主要邏輯入口, 這裡要主要channelName每次都是新物件, 但內部覆寫hashCode+equals。
    subscribe(codec, new ChannelName(channelName), promise, PubSubType.SUBSCRIBE, semaphore, listeners);
    return promise;
}

private void subscribe(Codec codec, ChannelName channelName,  RPromise<PubSubConnectionEntry> promise, PubSubType type, AsyncSemaphore lock, RedisPubSubListener<?>... listeners) {

    PubSubConnectionEntry connEntry = name2PubSubConnection.get(channelName);
    if (connEntry != null) {
        // 從已有Connection中取,如果存在直接把listeners加入到PubSubConnectionEntry中
        addListeners(channelName, promise, type, lock, connEntry, listeners);
        return;
    }

    // 沒有時,才是最重要的邏輯
    freePubSubLock.acquire(new Runnable() {

        @Override
        public void run() {
            if (promise.isDone()) {
                lock.release();
                freePubSubLock.release();
                return;
            }

            // 從佇列中取頭部元素
            PubSubConnectionEntry freeEntry = freePubSubConnections.peek();
            if (freeEntry == null) {
                // 第一次肯定是沒有的需要建立
                connect(codec, channelName, promise, type, lock, listeners);
                return;
            }

            // 如果存在則嘗試獲取,如果remainFreeAmount小於0則丟擲異常終止了。
            int remainFreeAmount = freeEntry.tryAcquire();
            if (remainFreeAmount == -1) {
                throw new IllegalStateException();
            }

            PubSubConnectionEntry oldEntry = name2PubSubConnection.putIfAbsent(channelName, freeEntry);
            if (oldEntry != null) {
                freeEntry.release();
                freePubSubLock.release();

                addListeners(channelName, promise, type, lock, oldEntry, listeners);
                return;
            }

            // 如果remainFreeAmount=0, 則從佇列中移除
            if (remainFreeAmount == 0) {
                freePubSubConnections.poll();
            }
            freePubSubLock.release();

            // 增加監聽
            RFuture<Void> subscribeFuture = addListeners(channelName, promise, type, lock, freeEntry, listeners);

            ChannelFuture future;
            if (PubSubType.PSUBSCRIBE == type) {
                future = freeEntry.psubscribe(codec, channelName);
            } else {
                future = freeEntry.subscribe(codec, channelName);
            }

            future.addListener(new ChannelFutureListener() {
                @Override
                public void operationComplete(ChannelFuture future) throws Exception {
                    if (!future.isSuccess()) {
                        if (!promise.isDone()) {
                            subscribeFuture.cancel(false);
                        }
                        return;
                    }

                    connectionManager.newTimeout(new TimerTask() {
                        @Override
                        public void run(Timeout timeout) throws Exception {
                            subscribeFuture.cancel(false);
                        }
                    }, config.getTimeout(), TimeUnit.MILLISECONDS);
                }
            });
        }

    });
}


private void connect(Codec codec, ChannelName channelName, RPromise<PubSubConnectionEntry> promise, PubSubType type, AsyncSemaphore lock, RedisPubSubListener<?>... listeners) {
    // 根據channelName計算出slot獲取PubSubConnection
    int slot = connectionManager.calcSlot(channelName.getName());
    RFuture<RedisPubSubConnection> connFuture = nextPubSubConnection(slot);
    promise.onComplete((res, e) -> {
        if (e != null) {
            ((RPromise<RedisPubSubConnection>) connFuture).tryFailure(e);
        }
    });


    connFuture.onComplete((conn, e) -> {
        if (e != null) {
            freePubSubLock.release();
            lock.release();
            promise.tryFailure(e);
            return;
        }

        // 這裡會從配置中讀取subscriptionsPerConnection
        PubSubConnectionEntry entry = new PubSubConnectionEntry(conn, config.getSubscriptionsPerConnection());
        // 每獲取一次,subscriptionsPerConnection就會減直到為0
        int remainFreeAmount = entry.tryAcquire();

        // 如果舊的存在,則將現有的entry釋放,然後將listeners加入到oldEntry中
        PubSubConnectionEntry oldEntry = name2PubSubConnection.putIfAbsent(channelName, entry);
        if (oldEntry != null) {
            releaseSubscribeConnection(slot, entry);

            freePubSubLock.release();

            addListeners(channelName, promise, type, lock, oldEntry, listeners);
            return;
        }


        if (remainFreeAmount > 0) {
            // 加入到佇列中
            freePubSubConnections.add(entry);
        }
        freePubSubLock.release();

        RFuture<Void> subscribeFuture = addListeners(channelName, promise, type, lock, entry, listeners);

        // 這裡真正的進行訂閱(底層與redis互動)
        ChannelFuture future;
        if (PubSubType.PSUBSCRIBE == type) {
            future = entry.psubscribe(codec, channelName);
        } else {
            future = entry.subscribe(codec, channelName);
        }

        future.addListener(new ChannelFutureListener() {
            @Override
            public void operationComplete(ChannelFuture future) throws Exception {
                if (!future.isSuccess()) {
                    if (!promise.isDone()) {
                        subscribeFuture.cancel(false);
                    }
                    return;
                }

                connectionManager.newTimeout(new TimerTask() {
                    @Override
                    public void run(Timeout timeout) throws Exception {
                        subscribeFuture.cancel(false);
                    }
                }, config.getTimeout(), TimeUnit.MILLISECONDS);
            }
        });
    });
}

PubSubConnectionEntry#tryAcquire方法, subscriptionsPerConnection代表了每個連線的最大訂閱數。當tryAcqcurie的時候會減少這個數量:

 public int tryAcquire() {
    while (true) {
        int value = subscribedChannelsAmount.get();
        if (value == 0) {
            return -1;
        }

        if (subscribedChannelsAmount.compareAndSet(value, value - 1)) {
            return value - 1;
        }
    }
}

梳理上述邏輯:

  • 1、還是進行重複判斷, 根據channelName從name2PubSubConnection中獲取,看是否存在已經訂閱:PubSubConnectionEntry; 如果存在直接把新的listener加入到PubSubConnectionEntry。
  • 2、從佇列freePubSubConnections中取公用的PubSubConnectionEntry, 如果沒有就進入connect()方法
    • 2.1 會根據subscriptionsPerConnection建立PubSubConnectionEntry, 然後呼叫其tryAcquire()方法 - 每呼叫一次就會減1
    • 2.2 將新的PubSubConnectionEntry放入全域性的name2PubSubConnection, 方便後續重複使用;
    • 2.3 同時也將PubSubConnectionEntry放入佇列freePubSubConnections中。- remainFreeAmount > 0
    • 2.4 後面就是進行底層的subscribeaddListener
  • 3、如果已經存在PubSubConnectionEntry,則利用已有的PubSubConnectionEntry進行tryAcquire;
  • 4、如果remainFreeAmount < 0 會丟擲IllegalStateException異常;如果remainFreeAmount=0,則會將其從佇列中移除, 那麼後續請求會重新獲取一個可用的連線
  • 5、最後也是進行底層的subscribeaddListener

三 總結

根因: 從上面程式碼分析, 導致問題的根因是因為PublishSubscribeService 會使用公共佇列中的freePubSubConnections, 如果同一個key一次性請求超過subscriptionsPerConnection它的預設值5時,remainFreeAmount就可能出現-1的情況, 那麼就會導致commandExecutor.syncSubscription(future)中等待超時,也就丟擲如上異常Subscribe timeout: (7500ms). Increase 'subscriptionsPerConnection' and/or 'subscriptionConnectionPoolSize' parameters.

解決方法: 在初始化Redisson可以可指定這個配置項的值。

相關引數的解釋以及預設值請參考官網:https://github.com/redisson/redisson/wiki/2.-Configuration#23-common-settings

相關文章