Nicolas Capponi created HTTPCORE-659:
----------------------------------------
Summary: Race condition in IOSessionImpl when setting event
Key: HTTPCORE-659
URL: https://issues.apache.org/jira/browse/HTTPCORE-659
Project: HttpComponents HttpCore
Issue Type: Bug
Components: HttpCore NIO
Affects Versions: 5.0.3
Reporter: Nicolas Capponi
There is a race condition in class {{org.apache.hc.core5.reactor.IOSessionImpl
}}when setting an event through the {{setEvent(final int op)}} method: there is
a check on the session to ensure it is not closed but this check is performed
outside the lock that is later acquired to protect the call on the
{{SelectionKey}}. If the session is closed just after the check but before
getting the lock, then it will throw a {{CancelledKeyException}} when trying to
perform the call on the {{SelectionKey}}.
I have implemented an HTTP server that use the reactive library
({{httpcore5-reactive}} component) and execute the requests in a separate
thread pool. I was hit by this issue, here is a stack trace (with traces non
specific to HttpCore library removed):
{code:java}
CancelledKeyException at SelectionKeyImpl.java:71
SelectionKeyImpl.java:90
IOSessionImpl.java:159
InternalDataChannel.java:324
AbstractHttp1StreamDuplexer.java: 452
ServerHttp1StreamDuplexer.java:136
ServerHttp1StreamHandler.java:94
ReactiveDataProducer.java:108
ReactiveDataProducer.java:100
... traces from publisher implementation using Rx java library...
ReactiveServerExchangeHandler.java:91
ReactiveServerExchangeHandler.java:83
...{code}
I was able to reproduce it almost systematically with unit test from my
application (concurrency issues are very sensitive to time). If necessary I can
provide a sample {{HttpAsyncServer}} which allows me to reproduce it by
introducing a simple {{Thread.sleep(50)}} sequence in {{IOSessionImpl}} code.
There is a simple fix which solves the problem: simply move the check on close
status inside the lock-unlock block:
{code:java}
// lock.lock();
try {
if (isStatusClosed()) {
return;
}
this.key.interestOps(this.key.interestOps() | op);
} finally {
lock.unlock();
}
{code}
Note that 2 other methods of this class are subject to this problem:
{{setEventMask(final int newValue)}} and \{{ clearEvent(final int op) }}. The
fix is the same than for the first method.
I can submit a pull request with the fix(es) if you indicate me in which branch
I should submit it.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
