Here is the relevant comment from Andreas:
--------------------------------
I did find a way to change the JVM to workaround this problem:
By creating a destructor for the thread pointer TLS we can restore the
value after pthread has set it to NULL.
Then when the native code destructor is run the thread pointer is still
intact.
Restoring a value in a pthread TLS is explicitly supported according to
the man page for pthread_key_create, and it will call the destructor for
the restored value again.
One would have to keep some extra state to make sure the destructor is
only called twice, since a pthread implementation is allowed to call the
destructor infinite times as long as the value is restored.
On my system pthread calls the destructor a maximum of four times, so
the attached JVM patch was sufficient as a proof of concept.
————————————————
Thomas writes:
Hi Brian,
The next patches where less straightforward. When running java I was
getting a ton of messages like:
Thread 832744400 has exited with leftover thread-specific data after 4
destructor iterations
After doing a lot of digging and debugging on Linux, I found the code path
for Linux was identical for Freebsd and the cleanup destructor was being
executed 4 times just like Freebsd, the difference being that Freebsd would
print out this benign warning while Linux would just ignore it. The
problem is that all threads that are created and initialize TLS current
thread data, must clean them up by explicitly setting the TLS current
thread to null. I’ve come up with two approaches to accomplish this.
clean up TLS current thread at end of ::run functions similar to how it's
done in openjdk8.
http://brian.timestudybuddy.com/webrev/hotspot__clear_thread_current/webrev/
clear current thread before exiting java_start to avoid warnings from
leftover pthread_setspecific data
http://brian.timestudybuddy.com/webrev/hotspot__clear_thread_current_alt/webrev/
I do not think this is a real leak. From what I remember of how the glibc
implements TLS, setting the TLS slot value to NULL would not in itself
delete anything. In VM, this slot keeps the pointer to the current Thread*,
which is correctly deleted at the end of the thread (void
JavaThread::thread_main_inner()).
Digging further, I found the pthread key destructor
"restore_thread_pointer(void* p)" in threadLocalStorage_posix.cpp:
// Restore the thread pointer if the destructor is called. This is in case
// someone from JNI code sets up a destructor with pthread_key_create to run
// detachCurrentThread on thread death. Unless we restore the thread
pointer we
// will hang or crash. When detachCurrentThread is called the key will be
set
// to null and we will not be called again. If detachCurrentThread is never
// called we could loop forever depending on the pthread implementation.
extern "C" void restore_thread_pointer(void* p) {
ThreadLocalStorage::set_thread((Thread*) p);
}
So, it seems we even reset deliberately the thread pointer to a non-NULL
value. The comment claims that we reset the Thread* value in case there is
another user-provided destructor which runs afterwards and which
does detachCurrentThread () which would require Thread::current() to work.
But there a details I do not understand:
- At this point, should the Thread* object not already be deallocated, so
this would be a dangling pointer anyway?
- Also, according to Posix, this is unspecified. Doc on
pthread_setspecific() states: "Calling pthread_setspecific() from a
thread-specific data destructor routine may result either in lost storage
(after at least PTHREAD_DESTRUCTOR_ITERATIONS attempts at destruction) or
in an infinite loop."
- In jdk8, we did reset the slot value to NULL before Thread exit. So, in
this case detachCurrentThread() from a pthread_key destructor should not
have worked at all.
Could someone from Oracle maybe shed light on this?
Please see the following discussion and bug report:
http://mail.openjdk.java.net/pipermail/hotspot-runtime-dev/2014-February/010759.htmlNote I don't follow this list so please include me directly in any follow-ups if needed.
Thanks,
David