Hotspot crash compiling sun.invoke.util.ValueConversions::identity

[Vladimir Ivanov]

Eirik,

Thanks a lot for the reproducer!

With fastdebug binaries it fails with the following error:
    #  assert(false) failed: DEBUG MESSAGE: i2c adapter must return to 
an interpreter frame

I wasn't able to reproduce the problem neither with 8u40 nor 9.

It should be JDK-8056154 [1].

Do you do class redefinition in the agent?

Best regards,
Vladimir Ivanov

[1] https://bugs.openjdk.java.net/browse/JDK-8056154

On 3/2/15 10:22 PM, Eirik Bjørsnøs wrote:
> This seems to reproduce the issue reliably on 1.8.0_31 on my Mac OS X Maverics:
>
> # TERMINAL A:
>
> git clone https://github.com/WinterTechForum/skybar.git
> git clone https://github.com/WinterTechForum/skybar-demo.git
> cd skybar
> git checkout 7eef1b8
> ./gradlew shadowJar
> cd ../skybar-demo/
> mvn clean install
> java -Dskybar.includedPackage=com/skybar
> -javaagent:../skybar/build/libs/skybar-1.0-SNAPSHOT-all.jar -jar
> target/skybar-demo-1.0-SNAPSHOT-jetty-console.war --headless
>
> # Terminal B:
>
> curl "http://localhost:8080/message?[1-100000]"
>
> (Let run for a minute or two)
>
> # Terminal C:
> curl "http://localhost:4321/coverage.json?[1-100000]"
>
> On Mon, Mar 2, 2015 at 4:26 PM, Vladimir Ivanov
> <vladimir.x.ivanov at oracle.com> wrote:
>> Thanks, Eirik!
>>
>>> I have an hs_err file ready with a corresponding core dump file (176MB).
>>>
>>> I'll zip these up and send a download link to you off-list.
>>>
>>> Let me know if there is a better way to manage these files. Should we
>>> have a Jira issue to track progress?
>>
>> Yes, I plan to file one once I make sure it's not a duplicate.
>>
>> Best regards,
>> Vladimir Ivanov
>>
>>
>>>
>>> Thanks,
>>> Eirik.
>>>
>>> On Mon, Mar 2, 2015 at 4:01 PM, Vladimir Ivanov
>>> <vladimir.x.ivanov at oracle.com> wrote:
>>>>
>>>> Eirik,
>>>>
>>>> Thanks for the report!
>>>>
>>>> As a first step, I'd like to see corresponding hotspot crash log
>>>> (hs_err_pid*.log).
>>>>
>>>> It seems you stepped on a bug in invokedynamic implementation (lambdas
>>>> are
>>>> translated into indy calls on bytecode level) which happens when it is
>>>> initialized too early in VM lifecycle (e.g. java agent premain).
>>>>
>>>> I'd be grateful for a stand-alone test case (focus on premain method
>>>> logic)
>>>> or steps how to reproduce it myself.
>>>>
>>>> Best regards,
>>>> Vladimir Ivanov
>>>>
>>>> On 3/2/15 4:10 AM, Eirik Bjørsnøs wrote:
>>>>>
>>>>>
>>>>> Ladies and gentlemen,
>>>>>
>>>>> Last week, some of the world's brightest, most capable and all round
>>>>> good looking Java developers gathered in Crested Butte, Colorado for
>>>>> the Winter Tech Forum (WTF) 2015 [1].
>>>>>
>>>>> Wednesday was Hackathon Day, and a group of us decided to embark the
>>>>> adventure of creating a live Java code coverage engine called Skybar
>>>>> [2].
>>>>>
>>>>> We made good progress and by the afternoon we knew we'd have something
>>>>> good to show at the 8pm showcase. So we confidently headed over to
>>>>> Ryce Asian Bistro [3] at 17:30 for a nice team dinner.
>>>>>
>>>>> Luckily we went easy on the saké, because when we came back to
>>>>> finalise our demo, we found that Hotspot had crashed:
>>>>>
>>>>> # A fatal error has been detected by the Java Runtime Environment:
>>>>> #
>>>>> #  SIGBUS (0xa) at pc=0x000000012a6d61e6, pid=10738, tid=18691
>>>>> [..]
>>>>> Compiled method (c2)  152453 1542       4
>>>>> sun.invoke.util.ValueConversions::identity (2 bytes)
>>>>>
>>>>> Shock, horror! Our demo was just minutes away, would we be able to
>>>>> present? Our team immediately went into debugging mode:
>>>>>
>>>>> Could we reproduce without using our -javaagent? No. Was the crash
>>>>> consistently reproducible? Kind of, but we had to stress test our
>>>>> code. Was there a any consistency in the hs_err reports? Yes, it
>>>>> seemed to always involve lambda expressions in out agent code.
>>>>>
>>>>> The team decided we had to remove all lambdas from our agent code in
>>>>> the hope that would improve the situation. And it did! At 7:50 PM, we
>>>>> had a stable build and ran out to make it to the presentations.
>>>>>
>>>>> Skybar was very well received. Our demo worked. Nothing crashed.
>>>>> Applause and shouts of "Ship it!" was heard from the audience. It was
>>>>> good times. We headed over to the Princess Wine Bar [4] for some
>>>>> celebratory drinks.
>>>>>
>>>>> The team is continuing our work on Skybar. But we miss our lamdas, and
>>>>> we want them back! To achieve this, we need help from the Hotspot
>>>>> team. On our side, we are willing to  work with you try to make
>>>>> reproducible test cases and collect any other useful information.
>>>>>
>>>>> What would be the best next step in saving Lambas on Agents? We would
>>>>> love to help out fixing this. Should we file an issue somewhere and
>>>>> upload our hs_err file(s)?
>>>>>
>>>>> On behalf of the Skybar team -
>>>>> Eirik,
>>>>>
>>>>> [1] http://www.mindviewinc.com/Conferences/JavaPosseRoundup/
>>>>> [2] https://github.com/WinterTechForum/skybar
>>>>> [3] http://www.ryceasianbistro.com
>>>>> [4] https://www.facebook.com/princesswinebar
>>>>>
>>>>
>>