Hi Chris,
On 9/08/2013 8:36 PM, Chris Hegarty wrote:
Firstly, I think the memory leak issue should be moved forward
separately to this cleanup effort. They are unrelated, and I'm starting
to get the feeling that this could take some time to reach conclusion.
It seems reasonable to separate the issues.
I agree. I'm sure when Alan suggested to check the return he didn't
expect it to unravel like this :) As we know hotspot will never
actually return NULL there is no urgency to add this in.
On 09/08/2013 10:27, Ivan Gerasimov wrote:
Chris,
I would use this
if ((name_utf = (*env)->GetStringUTFChars(env, name, &isCopy)) ==
NULL) {
JNU_ThrowOutOfMemoryError(env, "GetStringUTFChars failed");
return NULL/JNI_False/-1;
}
If I understand it correctly, JNU_ThrowOutOfMemoryError throws an
exception only if it hasn't been already thrown.
JNU_ThrowOutOfMemoryError is simply a convenience wrapper for
JNU_ThrowByName(env, "java/lang/OutOfMemoryError", msg);
...and JNU_ThrowByName [1] is defined as...
JNU_ThrowByName(JNIEnv *env, const char *name, const char *msg) {
class cls = (*env)->FindClass(env, name);
if (cls != 0) /* Otherwise an exception has already been thrown */
(*env)->ThrowNew(env, cls, msg);
}
}
Neither FindClass or ThrowNew is safe to call if there is a pending
exception [1].
Right - we have to check for a pending exception before trying to
throw one.
Now the issue comes down to; could there ever be a pending exception if
GetStringUTFChars returns NULL? The latest specs doesn't indicate that
there could be, but every copy of "The Java Native Interface
Programmer's Guide and Specification" I can find does. There also
appears to be an assumption of this if you look at the usages in the
JDK.
AFAIK there is only one version of the JNI spec book and it never got
updated. The official spec says no throw, but when people have the
book on their bookshelf that is what they tend to rely on. I looked at
some of the usages and they seem exception agnostic - many of them
don't even check for NULL :(
The implementation as it stands will not throw and will not return NULL.
I would really like to get a definitive answer on the JNI specification
for GetStringUTFChars before making any changes here.
The JNI spec (as opposed to the book) is definitive. If we don't like
what is there then it requires a spec change.
I can't find any reference to this particular issue being raised before.
Cheers,
David
-Chris.
[1]
http://hg.openjdk.java.net/jdk8/tl/jdk/file/54f0ccdd9ad7/src/share/native/common/jni_util.c
[2]
http://docs.oracle.com/javase/6/docs/technotes/guides/jni/spec/design.html#wp17626
Sincerely yours,
Ivan
On 09.08.2013 11:25, Chris Hegarty wrote:
On 09/08/2013 06:47, David Holmes wrote:
Sorry I messed this up. The JNI book says GetStringUTFChars will
return
NULL and post OOME but the JNI spec (latest version 6.0) does not
- it
only says it will return NULL on failure.
This is indeed strange. Most usages of this function in the jdk expect
the former. If this is not the case, then we may need to do an audit
of all usages.
So your previous version was the more correct. Given we just
failed to
allocate C-heap I think we are on thin ice anyway, but better to at
least attempt to do the right thing.
I'm not sure what the right thing to do here is? It seems a little
unwieldy!
if ((name_utf = (*env)->GetStringUTFChars(env, name, &isCopy)) ==
NULL) {
if ((*env)->ExceptionOccurred(env)) {
return NULL/JNI_False/-1;
} else {
throwException(env, "java/lang/InternalError", "GetStringUTFChars
failed");
return NULL/JNI_False/-1;
}
Given we have no idea why GetStringUTFChars may have failed, what
exception do we throw?
Also worth noting is that this bug fix has moved away from the
original problem (memory leak), and is now focused on code cleanup.
If we cannot get agreement on the cleanup, or it looks like more
clarification is needed around the cleanup effort, then I would like
to suggest that we proceed with the original fix for the memory leak
and separate out the cleanup effort.
-Chris.
FYI I filed 8022683 to fix GetStringUTFChars.
David
-----
On 9/08/2013 3:21 PM, David Holmes wrote:
Thumbs up!
Thanks,
David
On 9/08/2013 8:19 AM, Ivan Gerasimov wrote:
Thanks David!
On 09.08.2013 1:15, David Holmes wrote:
Main fix looks good to me.
Regression test may need some tweaking eg I think othervm will be
needed.
Yes, it's a good point.
Since there may be a memory leak in the test, it'd better not
interfere
with other tests in jtreg.
Also this:
System.out.println("WARNING: Cannot perform memory leak
detection on
this OS");
should probably just say something like "Test skipped on this
OS" -
there are other tests that do this so just check if there is
common
terminology. (In the future we may have keyword tags to flag
this as
linux only etc.)
OK, I changed the phrase to "Test only runs on Linux".
Updated webrev is here:
http://cr.openjdk.java.net/~igerasim/8022584/4/webrev/
Updated export is at the same place:
http://cr.openjdk.java.net/~igerasim/2commit/8022584-jdk8-Memleak-in-NetworkInterface.patch
Sincerely yours,
Ivan
Thanks,
David
On 9/08/2013 4:41 AM, Ivan Gerasimov wrote:
Chris, if it's not too late, I'd like to include a regtest in the
fix.
Here's webrev that includes the test:
http://cr.openjdk.java.net/~igerasim/8022584/3/webrev/
The test gets past with the fixed jdk8 and fails with
jdk8-b101 and
jdk7
as expected.
Sincerely yours,
Ivan
On 08.08.2013 17:50, Chris Hegarty wrote:
Looks good to me. Thanks Ivan.
-Chris.
On 08/08/2013 02:33 PM, Ivan Gerasimov wrote:
Hello Chris!
Here's the update:
http://cr.openjdk.java.net/~igerasim/8022584/2/webrev/
Thanks for offering the sponsorship!
Here's the hg-export
http://cr.openjdk.java.net/~igerasim/2commit/8022584-jdk8-Memleak-in-NetworkInterface.patch
Sincerely yours,
Ivan
On 08.08.2013 12:43, Chris Hegarty wrote:
Thanks for taking this Ivan.
Can you please make the changes suggested by both David and
Alan (
simply return NULL/-1/JNI_FALSE, as appropriate, if
GetStringUTFChars
fails ( returns NULL ), then I will sponsor this change into
jdk8
for
you.
Please post an update webrev to cr.openjdk.java.net.
-Chris.
On 08/08/2013 06:01 AM, David Holmes wrote:
Ivan,
On 8/08/2013 2:05 PM, Ivan Gerasimov wrote:
David, Alan,
I added checking for NULL results and throwing
OOMException if
necessary.
You don't need to throw it yourself:
JNU_ThrowOutOfMemoryError(env, NULL);
Assuming a correct VM implementation if NULL is returned
then an
OOME
should already be pending and will be thrown as soon as your
native
code
returns to Java.
Thanks,
David
I've also added some spaces along the code to improve
indentation.
Would you please review the updated webrev?
http://washi.ru.oracle.com/~igerasim/webrevs/8022584/1/webrev/
Sincerely yours,
Ivan
On 08.08.2013 5:35, David Holmes wrote:
On 8/08/2013 11:20 AM, Ivan Gerasimov wrote:
Thanks Alan!
I've checked that and it turns out that GetStringUTFChars
cannot
return
NULL.
For allocation of the result string it calls
AllocateHeap()
with
the
default EXIT_OOM fail strategy.
Thus, in case of being unable to allocate memory it simply
stops
VM.
Ouch! That is a hotspot bug. GetStringUTFChars is
supposed to
throw
OutOfMemoryError if it has memory issues, not abort the VM!
I recommend that you check for NULL and/or a pending
exception.
David
Sincerely yours,
Ivan
On 08.08.2013 5:05, Alan Bateman wrote:
(cc'ing net-dev).
The change looks okay to me. One suggestion (while you
are
there) is
to check the return from GetStringUTFChars so that the
name
returns
when it fails with NULL.
-Alan.
On 07/08/2013 17:46, Ivan Gerasimov wrote:
Hello everybody!
Some methods of NetworkInterface class were reported to
leak
memory.
http://bugs.sun.com/view_bug.do?bug_id=JDK-8022584 (not
yet
available.)
Examples are isUp() and isLoopback().
Affected versions of jdk are 6, 7 and 8
Would you please review a simple fix that removes the
unnecessary
allocation?
http://cr.openjdk.java.net/~igerasim/8022584/0/webrev/
Sincerely yours,
Ivan
