A further followup/clarification ...
On 21/03/2016 8:20 AM, David Holmes wrote:
Sorry I didn't seem this before my previous reply ...
On 21/03/2016 4:53 AM, Brian Gardner wrote:
I looked into the origination of other calls to set_thread(NULL) in
openjdk8.
*
hotspot/src/share/vm/gc_implementation/concurrentMarkSweep/concurrentMarkSweepThread.cpp:163:
ThreadLocalStorage::set_thread(NULL);
*
hotspot/src/share/vm/gc_implementation/shared/concurrentGCThread.cpp:91:
ThreadLocalStorage::set_thread(NULL);
* hotspot/src/share/vm/runtime/thread.cpp:1351:
ThreadLocalStorage::set_thread(NULL);
These are use cases where set_thread(NULL) was called roughly towards
the end of the threads run() method. Annotating them showed they’ve
been around prior to being open sourced. At some point in time there
was an effort made to ensure set_thread(NULL) was always called on all
threads, and there is a degradation in openjdk9 in these regards.
Any thread that terminates explicitly will do this as part of the Thread
destructor. For other threads it didn't seem necessary given they would
terminate only when the process terminated.
Some non-JavaThreads can actually terminate but AFAICS fail to ever
destroy the associated Thread instance. That would seem to be a bug in
itself. If we did ensure the destructor was called on these threads then
that was also deal with the removal of the set_thread(NULL). If the
threads do not terminate however, and are still running when the VM
terminates, then there is no place to put the "missing"
set_thread(NULL). Note this applies to Java threads to - any Java thread
still running at VM termination never has set_thread(NULL) called.
So I'm still somewhat at a loss to understand when a "missing"
set_thread(NULL) can cause a problem at VM termination?
Thanks,
David
If cleaning things up in the Thread destructor (Thread::~Thread) worked
on all platforms, it would be a clean way to handle the cleanup. But
since this doesn’t get called on all platforms, this logic should have
been moved to JavaThread::run(). WatcherThread::run initializes it’s
TLS and cleans it up, why not have JavaThread::run() do the same? I
don’t like JavaThread::run leaving it up to all it’s implementations to
clean themselves up.
This is not a platform specific issue. Thread::~Thread gets called, or
not, for the same set of threads on all platforms. The VMThread, nor
those other threads, is not a JavaThread so JavaThread::run does not
come into it. And for JavaThreads we don't want to do this cleanup in
JavaThread::run because we may still need to refer to the current
thread, so it is done during Thread::~Thread when the thread has
"terminated" as far as the VM is concerned.
The problem at hand only affects threads that never terminate/detach.
They previously would call set_thread(NULL) (though why is unclear), but
after my changes they don't. But that can be rectified and as I said I
will file a bug to handle that.
Please note however I am heading out on vacation in two days so this is
unlikely to get fixed until I return in a couple of weeks.
Thanks,
David
Brian Gardner
On Mar 19, 2016, at 11:46 PM, Brian Gardner <open...@getsnappy.com
<mailto:open...@getsnappy.com>> wrote:
Yes, I was referring to the call to
ThreadLocalStorage::set_thread(NULL).
The problem I'm trying to resolve are the pthread destructors being
called repeadly. On linux this isn’t a problem because the destructor
is called 4 times then silently gives up. On FreeBSD the destructor
is called 4 times then prints a warning to stderr, which is a problem,
although it is harmless.
The message below from the original thread states there are three
scenarios threads fall into in regards to the initial commit. The
third scenario is the problem scenario I just mentioned and while it
is ok on Linux, it isn’t ok on Freebsd because of the warnings to
stderr.
http://mail.openjdk.java.net/pipermail/hotspot-runtime-dev/2014-February/010796.html
I didn’t articulate it very well but I was trying to say that calling
ThreadLocalStorage::set_thread(NULL) durning thread cleanup is the
proper way to prevent the destructor from being called repeatedly. I
actually can’t think of an alternate way to do this.
In openjdk8 all threads clean themselves by calling
ThreadLocalStorage::set_thread(NULL). I’m going to do some more
research to locate the change sets for these calls. Perhaps they are
isolated to bsd-port branch. I’ll let everyone know what I find.
Kind regards,
Brian Gardner
On Mar 18, 2016, at 2:57 PM, David Holmes <david.hol...@oracle.com
<mailto:david.hol...@oracle.com>> wrote:
On 19/03/2016 2:58 AM, Brian Gardner wrote:
That explains the destructor. Looking at the initial change set that
came out of this bug, we also see the first spot where we set the TLS
current thread is set to NULL
http://hg.openjdk.java.net/jdk9/jdk9/hotspot/rev/469835cd5494
Are you referring to:
+ // Thread destructor usually does this.
+ ThreadLocalStorage::set_thread(NULL);
? That's only there because the VMThread destructor is never called.
So I think it’s safe to say that setting TLS current thread to NULL is
the correct way to set the state to "destroying thread" and preventing
the destructor from looping indefinitely.
I'm not sure exactly what you mean.
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.
————————————————
We implemented the basic patch, we don't do anything to ensure it was
called at most twice. We expect all well behaving apps to detach
threads from the JVM before they terminate.
David
-----
On Mar 17, 2016, at 9:02 PM, David Holmes <david.hol...@oracle.com
<mailto:david.hol...@oracle.com>
<mailto:david.hol...@oracle.com>> wrote:
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.html
Note I don't follow this list so please include me directly in any
follow-ups if needed.
Thanks,
David
