RFR: JDK-8215550: Debugger does not work after reattach

[serguei.spitsyn at oracle.com]

Hi Gary,

It looks good to me.
Nice discovery!

Thanks,
Serguei


On 1/30/19 15:03, gary.adams at oracle.com wrote:
> Second reviewer or is it trivial enough for one?
>
> On 1/30/19 1:57 PM, Chris Plummer wrote:
>> This looks good.
>>
>> thanks,
>>
>> Chris
>>
>> On 1/30/19 8:24 AM, Gary Adams wrote:
>>> Following the trail from debugLoop_run, at the bottom of the loop
>>> there is a path through debugInit_reset that involves 
>>> eventHandler_reset
>>> and eventually eventHelper_reset. This seems like a better place to
>>> clear the flag back to original state.
>>>
>>>   Revised webrev: http://cr.openjdk.java.net/~gadams/8215550/webrev.01/
>>>
>>> On 1/29/19, 6:11 PM, Chris Plummer wrote:
>>>> Ok, so you can't do a "cont" because threads are not suspended. 
>>>> That means someone resumed them. When/where was this done?
>>>>
>>>> Regarding threadIDs changing, my guess is that debugLoop_run() is 
>>>> re-entered when the new connection is established. This will result 
>>>> in commonRef_reset() being called, which invalidates all reference 
>>>> IDs, including threadIDs. So the first time the agent needs to send 
>>>> a threadID to the debugger, it needs to create a new one.
>>>>
>>>> Chris
>>>>
>>>> On 1/29/19 1:52 PM, gary.adams at oracle.com wrote:
>>>>> Issuing a "cont" in the second session does not work, because
>>>>> the threads are not suspended.
>>>>>
>>>>> It's interesting that the thread ids have all changed in
>>>>> the reconnected session.
>>>>>
>>>>> ...
>>>>>
>>>>> main[1] threads
>>>>> Group system:
>>>>>   (java.lang.ref.Reference$ReferenceHandler)0x374 Reference 
>>>>> Handler running
>>>>>   (java.lang.ref.Finalizer$FinalizerThread)0x375 Finalizer cond. 
>>>>> waiting
>>>>>   (java.lang.Thread)0x376                         Signal 
>>>>> Dispatcher running
>>>>> Group main:
>>>>>   (java.lang.Thread)0x1 main              running (at breakpoint)
>>>>> Group InnocuousThreadGroup:
>>>>>   (jdk.internal.misc.InnocuousThread)0x377 Common-Cleaner cond. 
>>>>> waiting
>>>>> main[1] exit
>>>>> ...
>>>>> > threads
>>>>> Group system:
>>>>>   (java.lang.ref.Reference$ReferenceHandler)0x3b2 Reference 
>>>>> Handler running
>>>>>   (java.lang.ref.Finalizer$FinalizerThread)0x3b3 Finalizer cond. 
>>>>> waiting
>>>>>   (java.lang.Thread)0x3b4                         Signal 
>>>>> Dispatcher running
>>>>> Group main:
>>>>>   (java.lang.Thread)0x3b7 main              running
>>>>> Group InnocuousThreadGroup:
>>>>>   (jdk.internal.misc.InnocuousThread)0x3b8 Common-Cleaner cond. 
>>>>> waiting
>>>>> > cont
>>>>> > Nothing suspended.
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>
>>>>> On 1/29/19 2:27 PM, Chris Plummer wrote:
>>>>>> What is the state of the threads after the detach? If they are 
>>>>>> all automatically resumed by the agent, then I think the 
>>>>>> unblocking should be done by the same code that resumes the 
>>>>>> threads. If they are still suspended, then why would we want to 
>>>>>> unblock when the next connection comes in? It should be up to the 
>>>>>> debugger to detect that all threads are suspended and act 
>>>>>> accordingly.
>>>>>>
>>>>>> Also, what happens if after attaching again you issue a "cont" 
>>>>>> command?
>>>>>>
>>>>>> Chris
>>>>>>
>>>>>> On 1/29/19 9:55 AM, Gary Adams wrote:
>>>>>>> As far as I can tell, the quit and exit commands are only 
>>>>>>> handled locally
>>>>>>> on the debugger side of the connection. There is no packet sent 
>>>>>>> to the
>>>>>>> libjdwp agentlib.
>>>>>>>
>>>>>>> On 1/29/19, 12:17 PM, Chris Plummer wrote:
>>>>>>>> Hi Gary,
>>>>>>>>
>>>>>>>> When the "exit" or "quit" is done, aren't all threads resumed 
>>>>>>>> at that point, and shouldn't that result in the command loop 
>>>>>>>> being unblocked?
>>>>>>>>
>>>>>>>> thanks,
>>>>>>>>
>>>>>>>> Chris
>>>>>>>>
>>>>>>>> On 1/29/19 8:09 AM, Gary Adams wrote:
>>>>>>>>> Significant protections are put in place to protect the 
>>>>>>>>> commandLoop
>>>>>>>>> from multiple events that that have a suspend-all policy. The
>>>>>>>>> commandLoop uses a special block variable to ensure only
>>>>>>>>> a VirtualMachine or ThreadReference call to resume() will unblock
>>>>>>>>> the commandLoop. See
>>>>>>>>>
>>>>>>>>> src/jdk.jdwp.agent/share/native/libjdwp/eventHelper.c
>>>>>>>>>
>>>>>>>>> In this particular bug report, the user was stopped at a 
>>>>>>>>> breakpoint
>>>>>>>>> when they sent the "exit" command. The same effect can be 
>>>>>>>>> produced
>>>>>>>>> with a "quit" command or any killing of the debugger process.
>>>>>>>>>
>>>>>>>>> When the second session is started the commandLoop is still 
>>>>>>>>> blocked,
>>>>>>>>> so a new breakpoint will never be dequeued from the commandQueue.
>>>>>>>>>
>>>>>>>>> The proposed workaround will ensure the commandLoop is unblocked
>>>>>>>>> when a new session is started.
>>>>>>>>>
>>>>>>>>>   Issue: https://bugs.openjdk.java.net/browse/JDK-8215550
>>>>>>>>>   Webrev: http://cr.openjdk.java.net/~gadams/8215550/webrev.00/
>>>>>>>>>
>>>>>>>>> All testing has been done by manually replicating the reported
>>>>>>>>> command sequences. I'll see if an existing breakpoint test can be
>>>>>>>>> enhanced to cover this scenario.
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>
>>>>>>
>>>>>>
>>>>>
>>>>
>>>>
>>>
>>
>>
>