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:
- 獲取當前執行緒的執行緒id;
- tryAquire嘗試獲取鎖,並返回ttl
- 如果ttl為空,則結束流程;否則進入後續邏輯;
- this.subscribe(threadId)訂閱當前執行緒,返回一個RFuture;
- 如果在指定時間沒有監聽到,則會產生如上異常。
- 訂閱成功後, 通過while(true)迴圈,一直嘗試獲取鎖
- 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();
}
而subscribeService
在MasterSlaveConnectionManager
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
subscribe
和addListener
- 2.1 會根據subscriptionsPerConnection建立PubSubConnectionEntry, 然後呼叫其
- 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
subscribe
和addListener
;
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