Rushabh S Shah created HDFS-12667:
-------------------------------------
Summary: KMSClientProvider#ValueQueue does synchronous fetch of
edeks in background async thread.
Key: HDFS-12667
URL: https://issues.apache.org/jira/browse/HDFS-12667
Project: Hadoop HDFS
Issue Type: Bug
Components: encryption, kms
Affects Versions: 3.0.0-alpha4
Reporter: Rushabh S Shah
Assignee: Rushabh S Shah
There are couple of issues in KMSClientProvider#ValueQueue.
1.
{code:title=ValueQueue.java|borderStyle=solid}
private final LoadingCache<String, LinkedBlockingQueue<E>> keyQueues;
// Stripped rwlocks based on key name to synchronize the queue from
// the sync'ed rw-thread and the background async refill thread.
private final List<ReadWriteLock> lockArray =
new ArrayList<>(LOCK_ARRAY_SIZE);
{code}
It hashes the key name into 16 buckets.
In the code chunk below,
{code:title=ValueQueue.java|borderStyle=solid}
public List<E> getAtMost(String keyName, int num) throws IOException,
ExecutionException {
...
...
readLock(keyName);
E val = keyQueue.poll();
readUnlock(keyName);
...
}
private void submitRefillTask(final String keyName,
final Queue<E> keyQueue) throws InterruptedException {
...
...
writeLock(keyName); // It holds the write lock while the key is
being asynchronously fetched. So the read requests for all the keys that hashes
to this bucket will essentially be blocked.
try {
if (keyQueue.size() < threshold && !isCanceled()) {
refiller.fillQueueForKey(name, keyQueue,
cacheSize - keyQueue.size());
}
...
} finally {
writeUnlock(keyName);
}
}
}
{code}
According to above code chunk, if two keys (lets say key1 and key2) hashes to
the same bucket (between 1 and 16), then if key1 is asynchronously being
refetched then all the getKey for key2 will be blocked.
2. Due to stripped rw locks, the asynchronous behavior of refill keys is now
synchronous to other handler threads.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]