[
https://issues.apache.org/jira/browse/HDFS-14088?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16694217#comment-16694217
]
Yuxuan Wang commented on HDFS-14088:
------------------------------------
Thanks Íñigo Goiri for reviewing.
Here's a new patch HDFS-14088.002.patch for trunk.
I'm sorry about not explaining clearly. The patch avoid synchronize the whole
thing with double check locking.
The original question is, after calling performFailover() and getProxy(), the
former RequestHedgingInvocationHandler instances should be deprecated can still
access currentUsedProxy. It can cause the handler success pass the if-condition
{code:java}
currentUsedProxy != null
{code}
but immediately it turn to null because of performFailover() called by
somebody.
My idea is, let the RequestHedgingInvocationHandler hold the currentUsedProxy
and RequestHedgingProxyProvider hold the currentUsedHandler which warp the
RequestHedgingInvocationHandler into a proxy like before. Fail over will set
currentUsedHandler to null and getProxy() will assign a new
RequestHedgingInvocationHandler to it and return. So we can avoid the
deprecated handler can still access null currentUsedProxy after calling
performFailover().
For the unit test, I add some metric check. The test's idea is mock a call and
sleep, then call performFailover(). The original code
{code:java}
if (currentUsedProxy != null) {
try {
Object retVal = method.invoke(currentUsedProxy.proxy, args);
LOG.debug("Invocation successful on [{}]",
currentUsedProxy.proxyInfo);
return retVal;
}
{code}
debug log can throw a NullPointerException.
I know the idea for unit test is a little tricky, but I can't figure out better
one.
Reformat the code.
> RequestHedgingProxyProvider can throw NullPointerException when failover due
> to no lock on currentUsedProxy
> -----------------------------------------------------------------------------------------------------------
>
> Key: HDFS-14088
> URL: https://issues.apache.org/jira/browse/HDFS-14088
> Project: Hadoop HDFS
> Issue Type: Bug
> Components: hdfs-client
> Reporter: Yuxuan Wang
> Assignee: Yuxuan Wang
> Priority: Major
> Attachments: HDFS-14088.001.patch, HDFS-14088.002.patch
>
>
> {code:java}
> if (currentUsedProxy != null) {
> try {
> Object retVal = method.invoke(currentUsedProxy.proxy, args);
> LOG.debug("Invocation successful on [{}]",
> currentUsedProxy.proxyInfo);
> {code}
> If a thread run try block and then other thread trigger a fail over calling
> method
> {code:java}
> @Override
> public synchronized void performFailover(T currentProxy) {
> toIgnore = this.currentUsedProxy.proxyInfo;
> this.currentUsedProxy = null;
> }
> {code}
> It will set currentUsedProxy to null, and the first thread can throw a
> NullPointerException.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
