[
https://issues.apache.org/jira/browse/DIRMINA-1107?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16843262#comment-16843262
]
Jonathan Valliere commented on DIRMINA-1107:
--------------------------------------------
The issue with SSL is that the order messages are encoded MUST be the same
order they are written to the socket. The {{scheduledEvents}} atomic is
essentially useless as implemented because it relies increment/decrements
instead of CAS operations leading to inconsistent behaviors. The {{sslLock}} is
also ONLY used in {{flushScheduledEvents}} which means it could be removed
without negatively impacting other operations. In order to ensure that
messages are pushed into the {{IoSession}} in the same order as they were
encoded within the mutex on {{sslHandler}}. It is my recommendation to break
{{flushScheudledEvents}} into flush operations for receive and write. Then
ensure that all {{flushFilterWrite}} occurs within the mutex on {{sslHandler}}
to ensure correct write order. This will cause write operations to be within
the mutex longer than before but does not require any additional locks since
all writing threads must obtain the write mutex anyway.
This patch is pushed to
[branch/DIRMINA-1107|https://gitbox.apache.org/repos/asf?p=mina.git;a=shortlog;h=refs/heads/DIRMINA-1107]
> SslHandler flushScheduledEvents race condition, redux
> -----------------------------------------------------
>
> Key: DIRMINA-1107
> URL: https://issues.apache.org/jira/browse/DIRMINA-1107
> Project: MINA
> Issue Type: Bug
> Affects Versions: 2.1.2
> Reporter: Guus der Kinderen
> Priority: Major
> Fix For: 2.1.3
>
>
> DIRMINA-1019 addresses a race condition in SslHandler, but unintentionally
> replaces it with another multithreading issue.
> The fix for DIRMINA-1019 introduces a counter that contains the number of
> events to be processed. A simplified version of the code is included below.
> {code:java}
> private final AtomicInteger scheduledEvents = new AtomicInteger(0);
> void flushScheduledEvents() {
> scheduledEvents.incrementAndGet();
> if (sslLock.tryLock()) {
> try {
> do {
> while ((event = filterWriteEventQueue.poll()) != null) {
> // ...
> }
>
> while ((event = messageReceivedEventQueue.poll()) != null){
> // ...
> }
> } while (scheduledEvents.decrementAndGet() > 0);
> } finally {
> sslLock.unlock();
> }
> }
> }{code}
> We have observed occasions where the value of {{scheduledEvents}} becomes a
> negative value, while at the same time {{filterWriteEventQueue}} go
> unprocessed.
> We suspect that this issue is triggered by a concurrency issue caused by the
> first thread decrementing the counter after a second thread incremented it,
> but before it attempted to acquire the lock.
> This allows the the first thread to empty the queues, decrementing the
> counter to zero and release the lock, after which the second thread acquires
> the lock successfully. Now, the second thread processes any elements in
> {{filterWriteEventQueue}}, and then processes any elements in
> {{messageReceivedEventQueue}}. If in between these two checks yet another
> thread adds a new element to {{filterWriteEventQueue}}, this element can go
> unprocessed (as the second thread does not loop, since the counter is zero or
> negative, and the third thread can fail to acquire the lock).
> It's a seemingly unlikely scenario, but we are observing the behavior when
> our systems are under high load.
> We've applied a code change after which this problem is no longer observed.
> We've removed the counter, and check on the size of the queues instead:
> {code:java}
> void flushScheduledEvents() {
> if (sslLock.tryLock()) {
> try {
> do {
> while ((event = filterWriteEventQueue.poll()) != null) {
> // ...
> }
>
> while ((event = messageReceivedEventQueue.poll()) != null){
> // ...
> }
> } while (!filterWriteEventQueue.isEmpty() ||
> !messageReceivedEventQueue.isEmpty());
> } finally {
> sslLock.unlock();
> }
> }
> }{code}
> This code change, as illustrated above, does introduce a new potential
> problem. Theoretically, an event could be added to the queues and
> {{flushScheduledEvents}} be called returning {{false}} for
> {{sslLock.tryLock()}}, exactly after another thread just finished the
> {{while}} loop, but before releasing the lock. This again would cause events
> to go unprocessed.
> We've not observed this problem in the wild yet, but we're uncomfortable
> applying this change as-is.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
