RFR(S) 8177901: JDWP exit error JVMTI_ERROR_WRONG_PHASE(112): on checking for an interface

serguei.spitsyn at oracle.com serguei.spitsyn at oracle.com
Fri Sep 15 06:52:39 UTC 2017


On 9/14/17 20:28, Daniel D. Daugherty wrote:
> On 9/14/17 9:12 PM, serguei.spitsyn at oracle.com wrote:
>> On 9/14/17 19:07, David Holmes wrote:
>>> On 15/09/2017 9:26 AM, serguei.spitsyn at oracle.com wrote:
>>>> Hi Dan,
>>>>
>>>> Thank you a lot for reviewing!
>>>> All your suggestions are accepted.
>>>>
>>>> The updated webrev:
>>>> http://cr.openjdk.java.net/~sspitsyn/webrevs/2017/hotspot/8177901-jdi-wrong_phase.2/ 
>>>
>>>
>>>
>>>
>>> Not sure the new flag was actually needed if the only way to get to 
>>> the commandLoop_exitVmDeathLockOnError in the event helper thread is 
>>> via completeCommand(command) - which is within the locked region. 
>>> But its harmless.
>>
>> Yes, it is harmless.
>
> Think of it as insurance against someone changing the code in
> the future. :-)

Yes, it is the point. :)

Thanks,
Serguei


>
> Dan
>
>
>>
>>
>>>
>>> So thumbs up.
>>
>> Thank you a lot, David!
>> Serguei
>>
>>>
>>> Cheers,
>>> David
>>>
>>>>
>>>>
>>>> Just a sanity check is needed.
>>>>
>>>> Thanks,
>>>> Serguei
>>>>
>>>>
>>>>
>>>> On 9/14/17 14:53, Daniel D. Daugherty wrote:
>>>>> On 9/4/17 11:23 PM, serguei.spitsyn at oracle.com wrote:
>>>>>> Hi David,
>>>>>>
>>>>>>
>>>>>> On 9/4/17 21:43, David Holmes wrote:
>>>>>>> Hi Serguei,
>>>>>>>
>>>>>>> This looks good to me. One not below.
>>>>>>>
>>>>>>> On 2/09/2017 6:30 PM, serguei.spitsyn at oracle.com wrote:
>>>>>>>> Please, review a fix for:
>>>>>>>> https://bugs.openjdk.java.net/browse/JDK-8177901
>>>>>>>>
>>>>>>>> Webrev:
>>>>>>>> http://cr.openjdk.java.net/~sspitsyn/webrevs/2017/hotspot/8177901-jdi-wrong_phase.1 
>>>>>>>>
>>>>>>>
>>>>>>> src/jdk.jdwp.agent/share/native/libjdwp/debugInit.c
>>>>>>>
>>>>>>> +     // Release commandLoop vmDeathLock if necessary
>>>>>>> +     void commandLoop_exitVmDeathLockOnError(void);
>>>>>>> +     commandLoop_exitVmDeathLockOnError();
>>>>>>>
>>>>>>> The declaration should be in the eventHelper.h header file. It 
>>>>>>> looks really odd to declare the function then call it.
>>>>>>
>>>>>> Ok, I'll move it to the header.
>>>>>
>>>>> I did look for a newer webrev and didn't find one.
>>>>>
>>>>>
>>>>> General comments:
>>>>>   - Please update the copyright years as needed before pushing.
>>>>>
>>>>> src/jdk.jdwp.agent/share/native/libjdwp/debugInit.c
>>>>>     No comments (other than agreeing with David H).
>>>>>
>>>>> src/jdk.jdwp.agent/share/native/libjdwp/eventHandler.c
>>>>>     L1289:     /*
>>>>>     L1290:      * The VM will die soon after the completion of 
>>>>> this callback - we
>>>>>     L1291:      * may need to do a final synchronization with the 
>>>>> command loop to
>>>>>     L1292:      * avoid the VM terminating with replying to the 
>>>>> final (resume)
>>>>>     L1293:      * command.
>>>>>     L1294:      */
>>>>>     L1295:     commandLoop_sync();
>>>>>
>>>>>         Seems like your addition of the call to commandLoop_sync()
>>>>>         might resolve what appears (to me) to be a placeholder
>>>>>         comment for a future change. Perhaps the comment should be
>>>>>         something like this:
>>>>>
>>>>>         /*
>>>>>          * The VM will die soon after the completion of this 
>>>>> callback -
>>>>>          * we synchronize with both the command loop and the debug 
>>>>> loop
>>>>>          * for a more orderly shutdown.
>>>>>          */
>>>>>
>>>>> src/jdk.jdwp.agent/share/native/libjdwp/eventHelper.c
>>>>>     L772: void commandLoop_exitVmDeathLockOnError() {
>>>>>         The '{' should be on a line by itself to keep in the
>>>>>         same existing style.
>>>>>
>>>>>     L127: static jrawMonitorID vmDeathLock;
>>>>>     L708:             debugMonitorEnter(vmDeathLock);
>>>>>     L714:             debugMonitorExit(vmDeathLock);
>>>>>     L778:     err = JVMTI_FUNC_PTR(gdata->jvmti, GetCurrentThread)
>>>>>     L779:               (gdata->jvmti, &cur_thread);
>>>>>     L794:     debugMonitorExit(vmDeathLock);
>>>>>         L794 assumes that the vmDeathLock is held by the Command
>>>>>         Loop Thread. Perhaps add a new volatile flag after L127
>>>>>         that is set to true after L708 and set to false before
>>>>>         L714. Check the flag before calling GetCurrentThread()
>>>>>         on L778 and return early if the flag is not true.
>>>>>
>>>>> src/jdk.jdwp.agent/share/native/libjdwp/eventHelper.h
>>>>>     L57: void commandLoop_sync(void); /* commandLoop sync with 
>>>>> cbVMDeath */
>>>>>         Extra space before 'cbVMDeath'.
>>>>>
>>>>> Dan
>>>>>
>>>>>
>>>>>>
>>>>>> Thank you a lot for he review!
>>>>>> Serguei
>>>>>>
>>>>>>>
>>>>>>> Thanks,
>>>>>>> David
>>>>>>> -----
>>>>>>>
>>>>>>>>
>>>>>>>> Summary:
>>>>>>>>
>>>>>>>>    The approach to fix this is to introduce a new lock 
>>>>>>>> vmDeathLock in eventHelper.c
>>>>>>>>    and use it to sync the commandLoop with the JVMTI event 
>>>>>>>> callback cbVMDeath
>>>>>>>>    the same way as it was done for debugLoop_run.
>>>>>>>>    (see the fix of: 
>>>>>>>> https://bugs.openjdk.java.net/browse/JDK-8134103)
>>>>>>>>
>>>>>>>>    One potential issue with it is that the commandLoop() 
>>>>>>>> transitively uses some helper
>>>>>>>>    functions (e.g. from util.c) that use the macro EXIT_ERROR, 
>>>>>>>> and so, can abort.
>>>>>>>>    It seems, in such a case the vmDeathLock will remain locked, 
>>>>>>>> and so,
>>>>>>>>    the cbVMDeath() will block on it causing a deadlock.
>>>>>>>>    Note, that these helper functions can be also called from 
>>>>>>>> different contexts/threads
>>>>>>>>    (not from the commandLoop thread only). In such contexts the 
>>>>>>>> commandLoop vmDeathLock
>>>>>>>>    is not being entered, and so, should not be exited.
>>>>>>>>
>>>>>>>>    To fix this potential issue, new function, 
>>>>>>>> commandLoop_exitVmDeathLockOnError(),
>>>>>>>>    is introduced, and it is called from the debugInit_exit().
>>>>>>>> The commandLoop_exitVmDeathLockOnError() always checks if 
>>>>>>>> current thread is
>>>>>>>>    the commandLoop thread and only in such a case unlocks the 
>>>>>>>> vmDeathLock.
>>>>>>>>
>>>>>>>> Testing:
>>>>>>>>    Ran the nsk.jdi, nsk.jdwp and jtreg jdk_jdi for both release 
>>>>>>>> and fastdebug builds.
>>>>>>>>    All tests are passed.
>>>>>>>>
>>>>>>>> Thanks,
>>>>>>>> Serguei
>>>>>>>>
>>>>>>
>>>>>
>>>>
>>
>



More information about the serviceability-dev mailing list