Details
-
Bug
-
Status: Closed
-
Major
-
Resolution: Fixed
-
None
-
None
Description
In a test I was running a JVM was told to exit and Geode's Shutdown Hook initiated cache shutdown. This thread hung once in a while either waiting for a reply to a release of a distributed lock or for a reply to a region-destroy message.
I traced this down by adding some logging to TCPConduit and DirectChannel and it's due to the changes for GEODE-7727 ("modify sender thread to detect relese of connection"). Those changes cause the P2P Handshake thread to stay active reading from a shared Connection socket. Unfortunately, when this thread eventually exits it is invoking removeEndpoint, which closes all the other connections to the other node. This is causing messages to be lost.
Here's an example:
one node (19919) fails to form a connection and invokes removeEndpoint for the other node (23898)
bridgegemfire1_19919/system.log: [info 2020/06/09 11:05:08.862 PDT <P2P handshake reader@31e492f-45> tid=0x14f] BRUCE: asyncClose closing Connection, uid=4 shared=true ordered=false remoteAddr=rs-Awesome-14-1023a0i3xlarge-hydra-client-8(bridgegemfire4_host1_23898:23898)<ec><v13>:41005 isReceiver=true java.lang.Exception: stack trace at org.apache.geode.internal.tcp.Connection.asyncClose(Connection.java:833) at org.apache.geode.internal.tcp.Connection.close(Connection.java:1338) at org.apache.geode.internal.tcp.Connection.closePartialConnect(Connection.java:1276) at org.apache.geode.internal.tcp.ConnectionTable.closeCon(ConnectionTable.java:612) at org.apache.geode.internal.tcp.ConnectionTable.closeCon(ConnectionTable.java:604) at org.apache.geode.internal.tcp.ConnectionTable.removeEndpoint(ConnectionTable.java:851) at org.apache.geode.internal.tcp.ConnectionTable.removeEndpoint(ConnectionTable.java:751) at org.apache.geode.internal.tcp.Connection.close(Connection.java:1400) at org.apache.geode.internal.tcp.Connection.requestClose(Connection.java:1268) at org.apache.geode.internal.tcp.Connection.readMessages(Connection.java:1661) at org.apache.geode.internal.tcp.Connection.run(Connection.java:1460) at java.base/java.lang.Thread.run(Thread.java:834)
The other node's shared/unordered connection was unexpectedly terminated, causing it to also invoke removeEndpoint():
bridgegemfire4_23898/system.log: [info 2020/06/09 11:05:08.862 PDT <P2P handshake reader@2e1ef958-4> tid=0x36] BRUCE: asyncClose closing Connection, uid=4 shared=true ordered=false remoteAddr=rs-Awesome-14-1023a0i3xlarge-hydra-client-8(bridgegemfire1_host1_19919:19919)<ec><v1>:41002 isReceiver=false java.lang.Exception: stack trace at org.apache.geode.internal.tcp.Connection.asyncClose(Connection.java:833) at org.apache.geode.internal.tcp.Connection.close(Connection.java:1338) at org.apache.geode.internal.tcp.Connection.requestClose(Connection.java:1268) at org.apache.geode.internal.tcp.Connection.readMessages(Connection.java:1619) at org.apache.geode.internal.tcp.Connection.run(Connection.java:1460) at java.base/java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1128) at java.base/java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:628) at java.base/java.lang.Thread.run(Thread.java:834) bridgegemfire4_23898/system.log: [info 2020/06/09 11:05:08.862 PDT <P2P handshake reader@2e1ef958-4> tid=0x36] BRUCE: removeEndpoint invoked for rs-Awesome-14-1023a0i3xlarge-hydra-client-8(bridgegemfire1_host1_19919:19919)<ec><v1>:41002 reason=SocketChannel.read returned EOF notifyDisconnect=true java.lang.Exception: stack trace at org.apache.geode.internal.tcp.ConnectionTable.removeEndpoint(ConnectionTable.java:758) at org.apache.geode.internal.tcp.ConnectionTable.removeEndpoint(ConnectionTable.java:751) at org.apache.geode.internal.tcp.Connection.close(Connection.java:1400) at org.apache.geode.internal.tcp.Connection.requestClose(Connection.java:1268) at org.apache.geode.internal.tcp.Connection.readMessages(Connection.java:1619) at org.apache.geode.internal.tcp.Connection.run(Connection.java:1460) at java.base/java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1128) at java.base/java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:628) at java.base/java.lang.Thread.run(Thread.java:834)
An unordered message like a Reply can be lost in this case if it is written to the Connection's socket but one of these background threads then closes the socket.
Here the logs show a reply being sent on a Connection with a UID (unique ID) of 69. The message is never received because, as you can see, the other node closes the socket for Connection with UID 69.
bridgegemfire4_23898/system.log: [info 2020/06/09 11:05:09.906 PDT <Pooled Waiting Message Processor 3> tid=0x60] BRUCE: sending DLockReleaseReplyMessage (serviceName=__PRLS; replyCode=OK; sender=rs-Awesome-14-1023a0i3xlarge-hydra-client-8(bridgegemfire4_host1_23898:23898)<ec><v13>:41005; processorId=212756) to [rs-Awesome-14-1023a0i3xlarge-hydra-client-8(bridgegemfire1_host1_19919:19919)<ec><v1>:41002(uid=69)] bridgegemfire1_19919/system.log: [info 2020/06/09 11:05:09.907 PDT <P2P handshake reader@2c6c307c-65> tid=0x1b0] BRUCE: asyncClose closing Connection, uid=69 shared=true ordered=false remoteAddr=rs-Awesome-14-1023a0i3xlarge-hydra-client-8(bridgegemfire4_host1_23898:23898)<ec><v13>:41005 isReceiver=true java.lang.Exception: stack trace at org.apache.geode.internal.tcp.Connection.asyncClose(Connection.java:833) at org.apache.geode.internal.tcp.Connection.close(Connection.java:1338) at org.apache.geode.internal.tcp.Connection.closePartialConnect(Connection.java:1276) at org.apache.geode.internal.tcp.ConnectionTable.closeCon(ConnectionTable.java:612) at org.apache.geode.internal.tcp.ConnectionTable.closeCon(ConnectionTable.java:604) at org.apache.geode.internal.tcp.ConnectionTable.removeEndpoint(ConnectionTable.java:851) at org.apache.geode.internal.tcp.ConnectionTable.removeEndpoint(ConnectionTable.java:751) at org.apache.geode.internal.tcp.Connection.close(Connection.java:1400) at org.apache.geode.internal.tcp.Connection.requestClose(Connection.java:1268) at org.apache.geode.internal.tcp.Connection.readMessages(Connection.java:1619) at org.apache.geode.internal.tcp.Connection.run(Connection.java:1460) at java.base/java.lang.Thread.run(Thread.java:834)
I don't think Connection termination should be invoking removeEndpoint at all. Endpoints should only be removed in response to membership changes.
Attachments
Issue Links
- links to