[
https://issues.apache.org/jira/browse/HTTPASYNC-86?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14249828#comment-14249828
]
Jose Dillet commented on HTTPASYNC-86:
--------------------------------------
Oleg,
I don't think the connection has been closed by any party. The problem is that
one thread in the client (in the I/O dispatcher) returns the connection to the
pool while other thread in the client (main thread) is still using the same
connection. So the main thread gets connection closed, but the connection is
still open AFAICT.
This is the relevant code:
DefaultClientExchangeHandlerImpl: 304 isOpen returns false in this case
if (!managedConn.isOpen()) {
failed(new ConnectionClosedException("Connection closed"));
}
org.apache.http.impl.nio.conn.CPoolProxy:144
public boolean isOpen() {
final CPoolEntry local = this.poolEntry;
if (local != null) {
return !local.isClosed();
} else {
return false; <--- false is returned here because poolEntry
was null (connection was returned to the pool)
}
}
I'll try to provide more evidence that the connection has not been closed by
inspecting the network traffic.
Jose
> Race condition in async client when a keep alive connection is reused in POST
> request method
> --------------------------------------------------------------------------------------------
>
> Key: HTTPASYNC-86
> URL: https://issues.apache.org/jira/browse/HTTPASYNC-86
> Project: HttpComponents HttpAsyncClient
> Issue Type: Bug
> Affects Versions: 4.0.2, 4.1-beta1
> Environment: Ubuntu 12.04 64 bits Kernel 3.2.0-74
> Reporter: Jose Dillet
> Attachments: AsyncClientRaceConditionSample.java
>
>
> I have found a race condition in the async client when a POST request reuses
> a keep alive connection. We have around 2%-4% of our runs in a Continuous
> Integration environment failing because of this issue.
> I have found a consistent way of forcing the failure by using a debugger, so
> I can force the events to happen in the order which uncovers the problem.
> I have tested it on httpasyncclient 4.0.2 / httpcore 4.3.2 and
> httpasyncclient 4.1-beta1 / httpcore 4.4-beta1 and it happens in both
> versions (the lines provided in the steps to reproduce are from version
> 4.0.2).
> How to reproduce:
> 1) Add these breakpoints:
> a - DefaultClientExchangeHandlerImpl:304
> b - DefaultClientExchangeHandlerImpl:168
> 2) Run AsyncClientRaceConditionSample sample program
> Processing first request:
> - Program hits breakpoint a -> Resume
> - Program hits breakpoint b -> Resume
> Processing second request:
> - Program hits breakpoints a and b (in different threads)
> - Resume a)
> - Resume b)
> (Probably adding a small sleep after DefaultClientExchangeHandlerImpl:303
> only for the testing would also uncover the issue).
> ConnectionClosedException is thrown with this stack trace:
> Exception in thread "main" java.util.concurrent.ExecutionException:
> org.apache.http.ConnectionClosedException: Connection closed
> at org.apache.http.concurrent.BasicFuture.getResult(BasicFuture.java:68)
> at org.apache.http.concurrent.BasicFuture.get(BasicFuture.java:77)
> at
> AsyncClientRaceConditionSample.process(AsyncClientRaceConditionSample.java:81)
> at
> AsyncClientRaceConditionSample.main(AsyncClientRaceConditionSample.java:45)
> Caused by: org.apache.http.ConnectionClosedException: Connection closed
> at
> org.apache.http.impl.nio.client.DefaultClientExchangeHandlerImpl.connectionAllocated(DefaultClientExchangeHandlerImpl.java:305)
> at
> org.apache.http.impl.nio.client.DefaultClientExchangeHandlerImpl.access$000(DefaultClientExchangeHandlerImpl.java:63)
> at
> org.apache.http.impl.nio.client.DefaultClientExchangeHandlerImpl$1.completed(DefaultClientExchangeHandlerImpl.java:360)
> at
> org.apache.http.impl.nio.client.DefaultClientExchangeHandlerImpl$1.completed(DefaultClientExchangeHandlerImpl.java:357)
> at
> org.apache.http.concurrent.BasicFuture.completed(BasicFuture.java:115)
> at
> org.apache.http.impl.nio.conn.PoolingNHttpClientConnectionManager$InternalPoolEntryCallback.completed(PoolingNHttpClientConnectionManager.java:465)
> at
> org.apache.http.impl.nio.conn.PoolingNHttpClientConnectionManager$InternalPoolEntryCallback.completed(PoolingNHttpClientConnectionManager.java:449)
> at
> org.apache.http.concurrent.BasicFuture.completed(BasicFuture.java:115)
> at
> org.apache.http.nio.pool.AbstractNIOConnPool.fireCallbacks(AbstractNIOConnPool.java:443)
> at
> org.apache.http.nio.pool.AbstractNIOConnPool.lease(AbstractNIOConnPool.java:276)
> at
> org.apache.http.impl.nio.conn.PoolingNHttpClientConnectionManager.requestConnection(PoolingNHttpClientConnectionManager.java:261)
> at
> org.apache.http.impl.nio.client.DefaultClientExchangeHandlerImpl.requestConnection(DefaultClientExchangeHandlerImpl.java:351)
> at
> org.apache.http.impl.nio.client.DefaultClientExchangeHandlerImpl.start(DefaultClientExchangeHandlerImpl.java:126)
> at
> org.apache.http.impl.nio.client.InternalHttpAsyncClient.execute(InternalHttpAsyncClient.java:129)
> at
> org.apache.http.impl.nio.client.CloseableHttpAsyncClient.execute(CloseableHttpAsyncClient.java:72)
> at
> org.apache.http.impl.nio.client.CloseableHttpAsyncClient.execute(CloseableHttpAsyncClient.java:102)
> at
> org.apache.http.impl.nio.client.CloseableHttpAsyncClient.execute(CloseableHttpAsyncClient.java:87)
> at
> AsyncClientRaceConditionSample.process(AsyncClientRaceConditionSample.java:80)
> ... 1 more
> Some analysis from my side:
> 1 - When the first request is processed, this is the execution order of
> relevant lines (and the thread where it is processed):
> a - (I/O dispatcher 1)
> DefaultClientExchangeHandlerImpl<T>.connectionAllocated(NHttpClientConnection)
> line: 303
> b - (I/O dispatcher 1)
> DefaultClientExchangeHandlerImpl<T>.connectionAllocated(NHttpClientConnection)
> line: 304
> c - (I/O dispatcher 1) CPoolProxy.detach() line: 60
> d - (I/O dispatcher 1)
> DefaultClientExchangeHandlerImpl<T>.responseCompleted() line: 168
> 2 - When the second request is processed c) might happen before b) which
> leads to the problem:
> a - (main)
> DefaultClientExchangeHandlerImpl<T>.connectionAllocated(NHttpClientConnection)
> line: 303
> c - (I/O dispatcher 1) CPoolProxy.detach() line: 60
> b - (main)
> DefaultClientExchangeHandlerImpl<T>.connectionAllocated(NHttpClientConnection)
> line: 304
> e - (main) BasicFuture<T>.failed(Exception) line: 122
> d - (I/O dispatcher 1)
> DefaultClientExchangeHandlerImpl<T>.responseCompleted() line: 168
> The problem is that in case 2 the connection can be closed before request is
> completed because a) is processed in main thread and c) is processed in I/O
> dispatcher thread, so the ordering b < c is not guaranteed.
> However in case 1 because all the events are processed in I/O dispatcher
> thread, the ordering b < c is guaranteed (this seems to be true even if there
> are several dispatcher threads).
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
