Hi!
I was curious to see what others have done in the past and took a look at about
15 different testcases for memory leaks in the jdk tree and basically found 3
patterns:
Pattern A: Use a loop that would exercise the relevant code for a finite number
of runs. If an OOME is not thrown, the test passes. In order for the test to
finish quickly, most people seem to specify a very small Xmx to limit the
number of iterations needed. (example:
java/util/concurrent/BlockingQueue/PollMemoryLeak.java)
Pattern B: Code similar to what I have submitted for this review that tries to
leak one object, explicitly triggers a GC event, and then confirms that the
WeakReference was nulled out. One thing to note is that in every case I found
of this the author chose to run multiple GC events, usually somewhere between 5
to 10000 iterations. (example: javax/management/Introspector/ClassLeakTest.java)
Pattern C: These cases used a WeakReference like Pattern B above, but instead of
depending on an explicit call to System.gc(), would intentionally fill the heap until an
OOME is thrown to trigger at least 1 "full" gc event. (example:
java/rmi/server/UnicastRemoteObject/unexportObject/UnexportLeak.java)
Patterns A and B were most common (with pattern B being slightly more common),
and I only found Pattern C 2 times
Strictly speaking, all 3 patters are problematic as they each depend on
undefined behavior in some way or another. Patterns A and C require the JVM to
handle OOMEs somewhat gracefully (Pattern A when we fail, Pattern C when we
pass), and pattern B requires an explicit call to System.gc to collect an
arbitrary object.
The fact that we do not routinely see false positives from these cases
indicates that in practice, they all seem to work OK for the time being.
After thinking about this for a bit, I actually really like pattern C. Even though the
JVM is not technically guaranteed to gracefully handle all OOME conditions, we try very
hard to be as robust as possible in the face of Java-heap OOME. Again, the fact that we
do not routinely see false positives from these tests proves that in practice, this works
very well. Even if an otherwise "correct" change (say in the JVM) resulted in
Pattern C suddenly failing, it is something we would almost certainly want to investigate
as it would indicate that the robustness of the JVM has somehow regressed.
I plan to modify my testcase to follow pattern C and resubmit for review. If
anyone has any other ideas or comments, I would be grateful for your input.
Cheers,
-Buck
On 2013/01/30 22:32, David Buck wrote:
Hi Alan!
Thank you for your help.
TL;DR version:
Strictly speaking (going only by the specifications), this could give us
false positives, but I believe this is VERY unlikely to actually happen
in real life.
Long version:
Yes, I gave this some thought myself. For example, on JRockit, if the
object were in old space and System.gc() only did a young collection
(the default behavior for JRockit), this test would result in a false
positive. In fact, as the JVM is allowed by the specification to
completely ignore explicit GC calls, we could never guarantee that we
would the WeakReference would always get nulled out.
That said, in pactice this works very well for both HotSpot and JRockit.
Every scenario I have tried it out on (with both JVMs) has provided the
expected result every single time (i.e. failing when expected; never
resulting in false positive otherwise). It seems that both of Oracle's
JVMs as currently implemented are very unlikely to run into any issues
here. Marking the test cases as "othervm" also helps to remove most edge
case scenarios where you could still somehow imagine this failing. (For
example, on a JRockit-like JVM, other tests running concurrently could
trigger a gc in the middle of this test resulting in the HashMap and its
contents being promoted to old space and the null reference not being
cleared during the call to System.gc() as expected.)
One option would be to mark this as a manually-run test if we wanted to
be extra cautious. What do you think?
Minor nit, should be WeakReference<Object> to avoid the raw type.
I will update the webrev once we have decided what (if anything) to do
regarding the risk of false positives.
Cheers,
-Buck
On 2013/01/30 22:09, Alan Bateman wrote:
On 29/01/2013 23:36, David Buck wrote:
Hi!
This is a review request to add only the test case for the following
OracleJDK issue :
[ 7042126 : (alt-rt) HashMap.clone implementation should be
re-examined ]
http://bugs.sun.com/bugdatabase/view_bug.do?bug_id=7042126
* please note: I just marked the bug as "public" this morning, so
there will be a short delay before the above link is available.
The issue (root cause) is not in OpenJDK (i.e. the problem was
OracleJDK specific), but the test case is valid for any Java SE
implementation so it should go into OpenJDK so we can prevent a
similar issue from ever happening in both releases moving forward. The
test case simply helps ensure that the contents of a HashMap are not
leaked when we clone the HashMap.
webrev:
[ Code Review for jdk ]
http://cr.openjdk.java.net/~dbuck/7042126/webrev.00/
How robust is this test? I'm concerned that it might fail intermittently
if the System.gc doesn't immediately GC the now un-references entries in
hm.
Minor nit, should be WeakReference<Object> to avoid the raw type.
-Alan.
