Hi Alex,
It seems to be hanged in check_socket_file() because AttachListener::deque()
could not return due to safepoint.
So it is good idea to use ThreadBlockInVM, but I think we can change as below
to reduce overhead:
```
{
ThreadBlockInVM tbivm(JavaThread::current());
while (AttachListener::transit_state(AL_INITIALIZING,
AL_NOT_INITIALIZED) !=
AL_NOT_INITIALIZED) {
os::naked_yield();
}
}
```
It might consume much CPU time with long safepoint, but I think it is corner
case because this issue occurs intermittently.
Thanks,
Yasumasa
On 2020/05/09 10:14, Alex Menkov wrote:
Hi all,
please review the fix for
https://bugs.openjdk.java.net/browse/JDK-8235211
webrev:
http://cr.openjdk.java.net/~amenkov/jdk15/RemovingUnixDomainSocket/webrev/
Test failures are caused by deadlock during attach listener restarting:
check_socket_file function aborts socket listening and waits while attach
listener state becomes AL_NOT_INITIALIZED (it happens when
AttachListener::dequeue returns NULL).
AttachListener::dequeue method is blocked in ThreadBlockInVM dtor.
To solve it ThreadBlockInVM was added inside waiting cycle in check_socket_file.
Other changes:
- made _listener (and _shutdown for aix) volatile as they are used by 2 threads
(attach listener thread and signal handler thread) without synchronization;
- added close() for listening socket on aix (before it had only shutdown() for
it);
- additional logging and some cleanup in the test;
- added handling of LingeredApp hang.
--alex
