SockJs heartbeat causes deadlock with XHR polling [SPR-14833]

[spring-projects-issues]

Grégory Coutant opened SPR-14833 and commented

When using XHR polling as a transport for SockJs, a deadlock can occur when one thread is sending a message at the same time the heartbeat run.

Stack from thread sending a message :

at org.springframework.web.socket.sockjs.transport.session.AbstractSockJsSession.cancelHeartbeat(AbstractSockJsSession.java:280) 
- waiting to lock [0x00000000820ce450] (a java.lang.Object) 
at org.springframework.web.socket.sockjs.transport.session.AbstractHttpSockJsSession.sendMessageInternal(AbstractHttpSockJsSession.java:301) 
- locked [0x00000000820cfce0] (a java.lang.Object) 
at org.springframework.web.socket.sockjs.transport.session.AbstractSockJsSession.sendMessage(AbstractSockJsSession.java:166)

Heartbeat task :

at org.springframework.web.socket.sockjs.transport.session.AbstractHttpSockJsSession.resetRequest(AbstractHttpSockJsSession.java:334) 
- waiting to lock [0x00000000820cfce0] (a java.lang.Object) 
at org.springframework.web.socket.sockjs.transport.session.PollingSockJsSession.writeFrame(PollingSockJsSession.java:83) 
at org.springframework.web.socket.sockjs.transport.session.AbstractSockJsSession.sendHeartbeat(AbstractSockJsSession.java:255) 
- locked [0x00000000820ce450] (a java.lang.Object) 
at org.springframework.web.socket.sockjs.transport.session.AbstractSockJsSession$HeartbeatTask.run(AbstractSockJsSession.java:451) 
- locked [0x00000000820ce450] (a java.lang.Object)

Looking at the code, there seem to be two locks : heartbeatLock in AbstractSockJsSession and responseLock in AbstractHttpSockJsSession. When calling sendMessage(), the lock order is responseLock then heartbeatLock. For the HeartbeatTask, the lock order is heartbeatLock then responseLock.

---

Affects: 4.2.8, 4.3.3

Issue Links:

• AbstractSockJsSession deadlock when tomcat WsSession is closing [SPR-14917] #19484 AbstractSockJsSession deadlock when tomcat WsSession is closing

Referenced from: commits 72e1f7e, 1c80d2a, 4ce0e6b

Backported to: 4.2.9

[spring-projects-issues]

Juergen Hoeller commented

Good point: We're using a shared response lock for all interactions of a session now. To be backported to 4.3.4 and 4.2.9.