Analysis on JDK-8022321 java/lang/ref/OOMEInReferenceHandler.java fails intermittently
Peter Levart
peter.levart at gmail.com
Sat Dec 21 16:50:11 UTC 2013
Hi David,
Is it possible to get the test output when it fails? It can fail in two
different ways. I can't look at the bug (not authorized)...
On 12/20/2013 10:54 AM, Chris Hegarty wrote:
> On 20 Dec 2013, at 04:33, Mandy Chung <mandy.chung at oracle.com> wrote:
>
>> Hi Srikalyan,
>>
>> Maybe you can get add an uncaught handler to see if you can get
>> any information.
> +1. With this, at least the next time we see this failure we should have a better idea where the OOM is coming from.
>
> -Chris.
We can try, but I think the VM already prints the stack-trace of the
exception by default and as far as I remember, OOME thrown by VM is
preallocated and does not contain a stack trace. So I suspect we'll see
nothing more with the suggested UEH.
Is it possible to include in test, a modified version of Reference class
that would be prepended to boot-classpath? For example, containing the
following ReferenceHandler:
private static class ReferenceHandler extends Thread {
ReferenceHandler(ThreadGroup g, String name) {
super(g, name);
}
private volatile int state;
@Override
public String toString() {
return super.toString() + "[state=" + state + "]";
}
public void run() {
for (;;) {
state = 1;
Reference<Object> r;
state = 2;
synchronized (lock) {
state = 3;
if (pending != null) {
state = 4;
r = pending;
state = 5;
pending = r.discovered;
state = 6;
r.discovered = null;
state = 7;
} else {
state = 8;
// The waiting on the lock may cause an OOME
because it may try to allocate
// exception objects, so also catch OOME here
to avoid silent exit of the
// reference handler thread.
//
// Explicitly define the order of the two
exceptions we catch here
// when waiting for the lock.
//
// We do not want to try to potentially load
the InterruptedException class
// (which would be done if this was its first
use, and InterruptedException
// were checked first) in this situation.
//
// This may lead to the VM not ever trying to
load the InterruptedException
// class again.
try {
state = 9;
try {
state = 10;
lock.wait();
state = 11;
} catch (InterruptedException x) { state =
12; }
state = 13;
} catch (OutOfMemoryError x) { state = 14; }
state = 15;
continue;
}
state = 16;
}
state = 17;
// Fast path for cleaners
if (r instanceof Cleaner) {
state = 18;
((Cleaner)r).clean();
state = 19;
continue;
}
state = 20;
ReferenceQueue<Object> q = (ReferenceQueue) r.queue;
state = 21;
if (q != ReferenceQueue.NULL) q.enqueue(r);
state = 22;
}
}
}
...then just include the toString of referenceHandlerThread instance as
part of the exception message at the end of the test:
...
...
// wait at most 10 seconds for success or failure
for (int i = 0; i < 20; i++) {
if (refQueue.poll() != null) {
// Reference Handler thread still working -> success
return;
}
System.gc();
Thread.sleep(500L); // wait a little to allow GC to do
it's work before allocating objects
if (!referenceHandlerThread.isAlive()) {
// Reference Handler thread died -> failure
throw new Exception("Reference Handler thread died.
referenceHandlerThread: " + referenceHandlerThread);
}
}
// no sure answer after 10 seconds
throw new IllegalStateException("Reference Handler thread
stuck. weakRef.get(): " + weakRef.get() +
", referenceHandlerThread: "
+ referenceHandlerThread);
}
This might be safer than using UEH since at the time the
UEH.uncaughtException() is called, the heap might still be full which
would prevent printing the message. The test makes sure the allocated
waste gets GCed before reporting the outcome...
I suspect that with the above, the failure message would print 8 <=
state <= 14 ...
When I was trying out the OOMEInReferenceHandler test, I experimented
with various arrangements of exception handlers in ReferenceHandler and
encountered one that in one ocassion allowed OOME to sneak-through or
re-induced it, but I haven't been able to explain why and also could not
reproduce it afterwards. So it might still be that something
interesting is happening between state 8 and 14.
Regards, Peter
>> I ran it for 1000 times but not able to duplicate
>> the failure. Did you run it with jtreg (I didn't)?
>>
>> Below is the patch to install a thread's uncaught handler that
>> you can take and try.
>>
>> diff --git a/test/java/lang/ref/OOMEInReferenceHandler.java b/test/java/lang/ref/OOMEInReferenceHand
>> ler.java
>> --- a/test/java/lang/ref/OOMEInReferenceHandler.java
>> +++ b/test/java/lang/ref/OOMEInReferenceHandler.java
>> @@ -51,6 +51,14 @@
>> return first;
>> }
>>
>> + static class UEH implements Thread.UncaughtExceptionHandler {
>> + public void uncaughtException(Thread t, Throwable e) {
>> + System.err.println("ERROR: " + t.getName() + " exception " +
>> + e.getMessage());
>> + e.printStackTrace();
>> + }
>> + }
>> +
>> public static void main(String[] args) throws Exception {
>> // preinitialize the InterruptedException class so that the reference handler
>> // does not die due to OOME when loading the class if it is the first use
>> @@ -77,6 +85,8 @@
>> throw new IllegalStateException("Couldn't find Reference Handler thread.");
>> }
>>
>> + referenceHandlerThread.setUncaughtExceptionHandler(new UEH());
>> +
>> ReferenceQueue<Object> refQueue = new ReferenceQueue<>();
>> Object referent = new Object();
>> WeakReference<Object> weakRef = new WeakReference<>(referent, refQueue);
>>
>> On 12/19/2013 6:57 PM, srikalyan chandrashekar wrote:
>>> Hi David Thanks for your comments, the unguarded part(clean and enqueue) in the Reference Handler thread does not seem to create any new objects, so it is the application(the test in this case) which is adding objects to heap and causing the Reference Handler to die with OOME. I am still unsure about the side effects of the code change and agree with your thoughts(on memory exhaustion test's reliability).
>>>
>>> PS: hotspot dev alias removed from CC.
>>>
>>> --
>>> Thanks
>>> kalyan
>>>
>>> On 12/19/13 5:08 PM, David Holmes wrote:
>>>> Hi Kalyan,
>>>>
>>>> This is not a hotspot issue so I'm moving this to core-libs, please drop hotspot from any replies.
>>>>
>>>> On 20/12/2013 6:26 AM, srikalyan wrote:
>>>>> Hi all, I have been working on the bug JDK-8022321
>>>>> <https://bugs.openjdk.java.net/browse/JDK-8022321> , this is a sporadic
>>>>> failure and the webrev is available here
>>>>> http://cr.openjdk.java.net/~srikchan/Regression/JDK-8022321_OOMEInReferenceHandler-webrev/
>>>> I'm really not sure what to make of this. We have a test that triggers an out-of-memory condition but the OOME can actually turn up in the ReferenceHandler thread causing it to terminate and the test to fail. We previously accounted for the non-obvious occurrences of OOME due to the Object.wait and the possible need to load the InterruptedException class - but still the OOME can appear where we don't want it. So finally you have just placed the whole for(;;) loop in a try/catch(OOME) that ignores the OOME. I'm certain that makes the test happy, but I'm not sure it is really what we want for the ReferenceHandler thread. If the OOME occurs while cleaning, or enqueuing then we will fail to clean and/or enqueue but there would be no indication that has occurred and I think that is a bigger problem than this test failing.
>>>>
>>>> There may be no way to make this test 100% reliable. In fact I'd suggest that no memory exhaustion test can be 100% reliable.
>>>>
>>>> David
>>>>
>>>>> *
>>>>> **"Root Cause:Still not known"*
>>>>> 2 places where there is a possibility for OOME
>>>>> 1) Cleaner.clean()
>>>>> 2) ReferenceQueue.enqueue()
>>>>>
>>>>> 1) The cleanup code in turn has 2 places where there is potential for
>>>>> throwing OOME,
>>>>> a) thunk Thread which is run from clean() method. This Runnable is
>>>>> passed to Cleaner and appears in the following classes
>>>>> java/nio/DirectByteBuffer.java
>>>>> sun/misc/Perf.java
>>>>> sun/nio/fs/NativeBuffer.java
>>>>> sun/nio/ch/IOVecWrapper.java
>>>>> sun/misc/Cleaner/ExitOnThrow.java
>>>>> However none of the above overridden implementations ever create an
>>>>> object in the clean() code.
>>>>> b) new PrivilegedAction created in try catch Exception block of
>>>>> clean() method but for this object to be created and to be held
>>>>> responsible for OOME an Exception(other than OOME) has to be thrown.
>>>>>
>>>>> 2) No new heap objects are created in the enqueue method nor anywhere in
>>>>> the deep call stack (VM.addFinalRefCount() etc) so this cannot be a
>>>>> potential cause.
>>>>>
>>>>> *Experimental change to java.lang.Reference.java* :
>>>>> - Put one more guard (try catch with OOME block) in the Reference
>>>>> Handler Thread which may give the Reference Handler a chance to cleanup.
>>>>> This is fixing the test failure (several 1000 runs with 0 failures)
>>>>> - Without the above change the test fails atleast 3-5 times for every
>>>>> 1000 run.
>>>>>
>>>>> *PS*: The code change is to a very critical part of JDK and i am fully
>>>>> not aware of the consequences of the change, hence seeking expert help
>>>>> here. Appreciate your time and inputs towards this.
>>>>>
More information about the core-libs-dev
mailing list