Pero Atanasov created AMQ-5493:
----------------------------------
Summary: KahaDB MessageDatabase race condition while stopping the
broker and cleaning up
Key: AMQ-5493
URL: https://issues.apache.org/jira/browse/AMQ-5493
Project: ActiveMQ
Issue Type: Bug
Components: Broker
Affects Versions: 5.10.0
Reporter: Pero Atanasov
This issue was seen on the surface via the following broker INFO log message:
2014-12-10 16:21:58,842 | INFO | KahaDB: Recovering checkpoint thread after
death | org.apache.activemq.store.kahadb.MessageDatabase | Thread-26
This means that the checkpoint thread is being revived unnecessarily while
stopping the broker. It could even happen that it is revived multiple times
before the closing process is completed.
To show the flow of the race condition, consider the following blocks of code
from
activemq/activemq-kahadb-store/src/main/java/org/apache/activemq/store/kahadb/MessageDatabase.java:
Lines 425 - 442 [in close()]
if( opened.compareAndSet(true, false)) {
... some code ...
if (metadata.page != null) {
checkpointUpdate(true);
}
... some code ...
}
Lines 1499 - 1501 [in checkpointUpdate(boolean)]
public void execute(Transaction tx) throws IOException {
checkpointUpdate(tx, cleanup);
}
Line 1524 [in checkpointUpdate(Transaction, boolean)]
metadata.ackMessageFileMapLocation = checkpointAckMessageFileMap();
Line 1735 [in checkpointAckMessageFileMap()]
Location location = store(new
KahaAckMessageFileMapCommand().setAckMessageFileMap(new
Buffer(baos.toByteArray())), nullCompletionCallback);
Lines 993 - 995 [in store(...)]
if (checkpointThread != null && !checkpointThread.isAlive()) {
startCheckpoint();
}
Lines 332 - 372 [In startCheckpoint()]
if (checkpointThread == null) {
start = true;
} else if (!checkpointThread.isAlive()) {
start = true;
LOG.info("KahaDB: Recovering checkpoint thread after death");
}
if (start) {
checkpointThread = new Thread("ActiveMQ Journal Checkpoint Worker") {
@Override
public void run() {
try {
... some code ...
while (opened.get()) {
... some code ...
}
} catch (InterruptedException e) {
// Looks like someone really wants us to exit this
thread...
} catch (IOException ioe) {
LOG.error("Checkpoint failed", ioe);
brokerService.handleIOException(ioe);
}
}
};
checkpointThread.setDaemon(true);
checkpointThread.start();
}
Here is the sequence of events, involving code above that shows the race:
1. in close() set "opened" to false
2. in startCheckpoint() "opened.get()" is false so checkpoint thread exits run
method and makes checkpointThread.isAlive() false
3. in store(...) calling startCheckpoint() which will revive the checkpoint
thread again
Added some INFO logs to confirm the order above:
2014-12-17 13:27:43,678 | INFO | patanasov: from unload() calling close() |
org.apache.activemq.store.kahadb.MessageDatabase | Thread-18
2014-12-17 13:27:43,678 | INFO | patanasov: close(): set opened to false;
calling checkpointUpdate(true) |
org.apache.activemq.store.kahadb.MessageDatabase | Thread-18
2014-12-17 13:27:43,678 | INFO | patanasov: startCheckpoint():
checkpointThread exiting its run method |
org.apache.activemq.store.kahadb.MessageDatabase | ActiveMQ Journal Checkpoint
Worker
2014-12-17 13:27:43,681 | INFO | patanasov: from store() calling
startCheckpoint() | org.apache.activemq.store.kahadb.MessageDatabase | Thread-18
2014-12-17 13:27:43,682 | INFO | KahaDB: Recovering checkpoint thread after
death | org.apache.activemq.store.kahadb.MessageDatabase | Thread-18
Based on my limited understanding of this code, this does not seem to have any
serious negative impacts, but it would be nice to be looked at by the
community.
It does not seem to make sense to revive the thread after opened.get() became
false because the checkpointThread will not do anything anyway due to
opened.get() being false at that point. Consider the body of the
checkpointThread:
MessageDatabase.java, Lines 348 - 359
while (opened.get()){
... code that will attempt updates ...
}
However, since opened.get() is false, this will not enter the while loop and
hence the checkpointThread will exit its run quickly again.
A possible fix for this is the following
Lines 993 - 995 [in store(...)]
change
if (checkpointThread != null && !checkpointThread.isAlive()) {
startCheckpoint();
}
to
if (checkpointThread != null && !checkpointThread.isAlive() && opened.get()) {
startCheckpoint();
}
If opened.get() is false, then we must be closing and we will not revive the
thread.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)