[fedora-java] [Crash-catcher] java reports by crash catcher

[Jiri Moskovcak]

On 10/11/2010 04:03 PM, Mark Wielaard wrote:
> Looks like the java-devel list disgarded any email from non-subscribers
> (so that included my initial email and your reply). Hope that is fixed
> now. I left all text in the email, even though my reply consists of only
> on tiny paragraph so others can catch up on the discussion.
>
> On Mon, 2010-10-11 at 15:32 +0200, Jiri Moskovcak wrote:
>> On 10/11/2010 03:19 PM, Mark Wielaard wrote:
>>> Hi Jiri,
>>>
>>> On Wed, 2010-10-06 at 10:11 +0200, Jiri Moskovcak wrote:
>>>> what about resurrecting this feature for F15? We made some changes to
>>>> abrt so it uses socket instead of the helper app, so we can send you the
>>>> info on how to use it if you'd be interested in implementing it to javavm.
>>>
>>> I currently don't have the time to work on this, but do think they are
>>> good ideas that would improve abrt a lot for java/jvm based packages.
>>>
>>> I added Fedora Java Developers list to the CC in the hope someone would
>>> be interested. So if someone is looking for a fun (python!) hack, this
>>> might be interesting.
>>>
>> Actually it would be probably more C fun.
>>
>>> One idea is that abrt/crash-catcher creates a lot of bugzilla reports
>>> against the jvm package. Those do include a native backtrace, but don't
>>> include the (often far more useful) hs_err_pid.log file. It would be
>>> nice if abrt would find it and offer to automagically attach it to the
>>> bug report.
>>
>> Attaching the file is not a problem, but does it live in some
>> predictable location?
>
> It lives in the current working directory of the JVM. In theory one
> could hack the hotspot sources to place it somewhere else. But what
> would be a good place?
>

ABRT knows the cwd and the pid of the crashing process so it should be 
able to read hss_err_<pid>.log file I will play with it.

>>> The other idea discussed was when a java program exits through an
>>> uncaught exception in the main thread. In that case you might want to
>>> catch that and create a bug report against the package that contains the
>>> main class file (instead of against the jvm package).
>>>
>>
>> This one is actually what I'd like to be done in F15.
>>
>>> The first idea is probably the least work and has the most benefit in
>>> the short term (at least for the java-1.6.0-openjdk bug maintainers).
>>>
>>>> On 11/16/2009 09:43 AM, Jiri Moskovcak wrote:
>>>>> On 11/13/2009 11:00 AM, Mark Wielaard wrote:
>>>>>> On Wed, 2009-11-11 at 11:05 +0100, Jiri Moskovcak wrote:
>>>>>>> On 11/05/2009 05:49 PM, Mark Wielaard wrote:
>>>>>>>> The code is already setup to save if in a different place if necessary
>>>>>>>> (in fact if the current directory isn't writable for the user it will
>>>>>>>> try saving in /tmp). If /var/log/java is made writable for all users
>>>>>>>> that could be a place to dump the log also.
>>>>>>>>
>>>>>>>> The code can also be modified to actually call abrt (or an helper
>>>>>>>> executable/script) if necessary either with the path of the log file or
>>>>>>>> even with an open file descriptor to the log.
>>>>>>>>
>>>>>>> We use a helper to handle the python logs, so I think we could use the
>>>>>>> same helper for saving the java logs.
>>>>>>
>>>>>> Do you have a pointer to the code for this helper?
>>>>>
>>>>> http://git.fedorahosted.org/git/abrt.git?p=abrt.git;a=blob;f=src/Hooks/abrt-pyhook-helper.cpp;h=348fbc72bd12b3b0f6757bac69a44e725706cf5f;hb=HEAD
>>>>>
>>>>>
>>>>>>
>>>>>>>>> The other thing I have in mind is how to catch an unhandled
>>>>>>>>> exception in
>>>>>>>>> java programs, because in this case the VM exits "normally" and abrt
>>>>>>>>> can't detect it. We managed to catch these exception in python by
>>>>>>>>> overriding the default exception handler by script that is
>>>>>>>>> automatically
>>>>>>>>> loaded everytime when python VM is started. If there would be some way
>>>>>>>>> to this for java we could wire this to ABRT.
>>>>>>>>
>>>>>>>> In principal we could install some uncaught exception handler, but
>>>>>>>> uncaught exceptions might not be fatal (although they are admittedly
>>>>>>>> sloppy). The program may even happily run even if one thread has an
>>>>>>>> uncaught exception (as long as there are other non-daemon threads
>>>>>>>> running).
>>>>>>>>
>>>>>>> ABRT doesn't care if it is or isn't fatal, this is up to your exception
>>>>>>> handler - you can just log the exception using the abrt helper (because
>>>>>>> even if the exception is not fatal, it's usually a bug..) and contiune
>>>>>>> running the program.
>>>>>>
>>>>>> This part is trickier than the above. In case of a JVM crash there is a
>>>>>> clear point where we catch that crash and produce the necessary logs for
>>>>>> a bug report. In case of an application specific uncaught exception
>>>>>> there is an uncaught exception handler mechanism, but the application
>>>>>> could already be using it (either for a specific Thread, the ThreadGroup
>>>>>> or system wide). This might require some surgery to get right (and
>>>>>> unobtrusive for the application running on the JVM).
>>>>>>
>>>>>
>>>>> This is really up to you as I don't know much about about insides of JVM.
>>>>>
>>>>>>>> How would you determine which package the exception belongs to? For
>>>>>>>> a VM
>>>>>>>> crash it is almost always the java VM package that should get the bug
>>>>>>>> report (since the VM just shouldn't crash ever).
>>>>>>>
>>>>>>> If I'm right the Java VM is compiled, so it creates a coredump and would
>>>>>>> be handled by a different hook then scripts, but that applies only if
>>>>>>> you don't catch the sigsegv, sigabrt (whichI think you do, to create the
>>>>>>> logs..) and let it die.
>>>>>>
>>>>>> Yes, the VM catches fatal signals and creates an hs_err log file based
>>>>>> on the information it can still retrieve at that point before dying.
>>>>>>
>>>>>
>>>>> Ok, so we need to find the way how to pass this logs to abrt hook.
>>>>>
>>>>>>> I think, as Java VM is a non-trivial programme,
>>>>>>> we should write a special handler for it, or we can try to improve the
>>>>>>> general hook for compiled programs to be able to handle some additional
>>>>>>> data as the log file if that would be enough.
>>>>>>>
>>>>>>>> But for uncaught
>>>>>>>> exceptions reporting it against the java VM package is definitely the
>>>>>>>> wrong thing to do. How do you solve that in the python case?
>>>>>>>>
>>>>>>>
>>>>>>> The python exception hook is run in the context of the running script,
>>>>>>> so it knows the script name and the path to the script and then we can
>>>>>>> simply run $ rpm -qf /path/to/script to determine the package, the code
>>>>>>> to do this is:
>>>>>>>
>>>>>>> executable = os.path.abspath(sys.argv[0])
>>>>>>
>>>>>> Aha. I think we could determine the main class that is being run and the
>>>>>> classpath with .jar/.zip files that this class comes from. With that we
>>>>>> could probably achieve similar heuristics about the package that
>>>>>> provided the classes.
>>>>>>
>>>>>> Thanks,
>>>>>>
>>>>>> Mark
>