Don't deregister Channel as part of closing it when using native kqueue transport (#8881)

Motivation:

In https://github.com/netty/netty/pull/8665 we changed how we handle the registration of Channels to KQueue but missed to removed some code which would deregister the Channel before it actual closed the underlying socket. This could lead to have events triggered still while not have a mapping to the Channel anymore.

Modifications:

Remove deregister call during socket closure.

Result:

Fixes https://github.com/netty/netty/issues/8849.
This commit is contained in:
Norman Maurer 2019-02-25 08:55:55 +01:00 committed by GitHub
parent f176384a72
commit d02b51965f
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23

View File

@ -124,35 +124,7 @@ abstract class AbstractKQueueChannel extends AbstractChannel implements UnixChan
// Even if we allow half closed sockets we should give up on reading. Otherwise we may allow a read attempt on a
// socket which has not even been connected yet. This has been observed to block during unit tests.
inputClosedSeenErrorOnRead = true;
try {
if (isRegistered()) {
// The FD will be closed, which should take care of deleting any associated events from kqueue, but
// since we rely upon jniSelfRef to be consistent we make sure that we clear this reference out for
// all events which are pending in kqueue to avoid referencing a deleted pointer at a later time.
// Need to check if we are on the EventLoop as doClose() may be triggered by the GlobalEventExecutor
// if SO_LINGER is used.
//
// See https://github.com/netty/netty/issues/7159
EventLoop loop = eventLoop();
if (loop.inEventLoop()) {
doDeregister();
} else {
loop.execute(new Runnable() {
@Override
public void run() {
try {
doDeregister();
} catch (Throwable cause) {
pipeline().fireExceptionCaught(cause);
}
}
});
}
}
} finally {
socket.close();
}
socket.close();
}
@Override