RFR 8150689: Thread dump report "waiting to re-lock in wait()" shows incorrectly
Daniel D. Daugherty
daniel.daugherty at oracle.com
Mon Nov 26 19:56:08 UTC 2018
On 11/26/18 2:21 PM, Patricio Chilano wrote:
> Hi Dan,
>
> I reverted the change on safepoint.cpp and just removed the extra
> "print_thread_state_on()" line on thread.cpp.
> Here is the new webrev:
>
> http://cr.openjdk.java.net/~pchilanomate/8150689.02/webrev
src/hotspot/share/runtime/thread.cpp
Okay. So the extra print_thread_state_on() call output that would
appear
before the "Thread: ..." output line from _safepoint_state->print_on()
is now gone. Good.
The print_thread_state_on() output will still appear after the
"Thread: ..."
line in a _safepoint_state->print_on() call which means that safepoint
logging will still have that info. That should address David's concern.
src/hotspot/share/runtime/vframe.cpp
No comments.
Thumbs up on the code change. Maybe I missed it, but it is not clear
from a quick scan of this email what kind of testing was done. I'm
strongly hoping that we don't have any tests that rely on either:
- print_thread_state_on() output happening before "Thread: ..."
- two print_thread_state_on() output line
> Is this reviewed?
Yes, but please wait for David H. to chime in and confirm also.
Dan
>
> Thanks!
> Patricio
>
> On 11/21/18 7:18 PM, Daniel D. Daugherty wrote:
>> On 11/21/18 5:45 PM, Patricio Chilano wrote:
>>> Hi Dan,
>>>
>>> On 11/21/18 2:52 PM, Daniel D. Daugherty wrote:
>>>> On 11/20/18 12:51 PM, Patricio Chilano wrote:
>>>>> Hi all,
>>>>>
>>>>> Could you review this small fix for thread dump reports?
>>>>> Now the "waiting to re-lock in wait()" message is shown in the "at
>>>>> java.lang.Object.wait(Native Method)" frame where the re-locking
>>>>> is actually occurring. In the frame where the lock was first taken
>>>>> we always show "locked". Please check if there is some scenario
>>>>> where you think the lock info on the stack is not being printed as
>>>>> you would expect.
>>>>>
>>>>> I also noticed the _thread_state attribute was printed twice so I
>>>>> removed one instance. I don't know if there was a specific reason
>>>>> to print it twice so I can add it back.
>>>>>
>>>>> Webrev URL:
>>>>> http://cr.openjdk.java.net/~pchilanomate/8150689.01/webrev
>>>>
>>>> src/hotspot/share/runtime/safepoint.cpp
>>>> new L1122: _thread->print_thread_state_on(st);
>>>> L1123: st->print_cr("Thread: " INTPTR_FORMAT
>>>> L1124: " [0x%2x] State: %s _has_called_back
>>>> %d _at_poll_safepoint %d",
>>>> L1125: p2i(_thread),
>>>> _thread->osthread()->thread_id(), s, _has_called_back,
>>>> L1126: _at_poll_safepoint);
>>>> old L1127: _thread->print_thread_state_on(st);
>>>>
>>>> The reason for moving the line isn't jumping out at me (yet).
>>>> Update: see next file's comments.
>>>>
>>>> src/hotspot/share/runtime/thread.cpp
>>>> L2945: #ifndef PRODUCT
>>>> old L2946: print_thread_state_on(st);
>>>> L2947: _safepoint_state->print_on(st);
>>>> L2948: #endif // PRODUCT
>>>>
>>>> Okay so now I understand the change to safepoint.cpp. I agree
>>>> with deleting L2946 from thread.cpp because it would result in
>>>> two print_thread_state_on() calls in non-PRODUCT bits.
>>>>
>>>> However, I don't agree with moving the print_thread_state_on()
>>>> call in safepoint.cpp. If you do that then you'll change the
>>>> output order for the "Thread: ..." line to be after the
>>>> print_thread_state_on() output for all build configs.
>>> We could keep it in the same order in
>>> ThreadSafepointState::print_on() and if JavaThreadState and
>>> ThreadStatus need to be printed together maybe we could add a flag
>>> to decide whether to print the JavaThreadState in
>>> ThreadSafepointState::print_on(). (?)
>>>
>>>> So I just took a look at the output for baseline release bits
>>>> on my MBP and I'm surprised to see this:
>>>>
>>>> "main" #1 prio=5 os_prio=31 cpu=121.30ms elapsed=1.41s
>>>> tid=0x00007f8abb808800 nid=0x2003 waiting on condition
>>>> [0x0000000100aff000]
>>>> java.lang.Thread.State: TIMED_WAITING (sleeping)
>>>> JavaThread state: _thread_blocked
>>>> Thread: 0x00007f8abb808800 [0x2003] State: _at_safepoint
>>>> _has_called_back 0 _at_poll_safepoint 0
>>>> JavaThread state: _thread_blocked
>>>> at java.lang.Thread.sleep(java.base at 12-internal/Native Method)
>>>> at Sleeper.main(Sleeper.java:5)
>>>>
>>>> So even in baseline 'release' bits, I'm seeing two of these:
>>>>
>>>> JavaThread state: _thread_blocked
>>>>
>>>> I don't know what to think here. Can you explain why release
>>>> bits are also showing double print_thread_state_on() output?
>>>> Maybe I'm not understanding about the thread.cpp code.
>>> Odd, I did the same test in my MBP for a "release" build but I don't
>>> see the extra output. I also tested it for Linux and it doesn't show
>>> the extra output either. Can you confirm this?
>>
>> I just reran my Sleeper.java and did a CTRL-\ thread dump on
>> both Linux and MacOS with release bits. This time I did not
>> get the extra output with either...
>>
>> I just scrolled back thru my window to see what I did... I
>> set JAVA_HOME to the release bits and then I pasted the full
>> path to the fastdebug bits on my MBP... So instead of using
>> $JAVA_HOME/bin/java Sleeper... I ran the wrong version...
>>
>> Sorry for the confusion.
>>
>> Dan
>>
>>
>>>
>>>> So I went back to JDK8 release bits on my Solaris box:
>>>>
>>>> "main" #1 prio=5 os_prio=64 tid=0x000000000041e800 nid=0x2 waiting
>>>> on condition [0xffff80ffbf1ae000]
>>>> java.lang.Thread.State: TIMED_WAITING (sleeping)
>>>> at java.lang.Thread.sleep(Native Method)
>>>> at Sleeper.main(Sleeper.java:5)
>>>>
>>>> So there's no "JavaThread state: _thread_blocked" line at all.
>>>> I'm guessing we added that to 'release' bits output after
>>>> JDK8.
>>>>
>>>> And here's the JDK8 fastdebug bits on my Solaris box:
>>>>
>>>> "main" #1 prio=5 os_prio=64 tid=0x0000000000421800 nid=0x2 waiting
>>>> on condition [0xffff80ffbf1ae000]
>>>> java.lang.Thread.State: TIMED_WAITING (sleeping)
>>>> JavaThread state: _thread_blocked
>>>> Thread: 0x0000000000421800 [0x 2] State: _at_safepoint
>>>> _has_called_back 0 _at_poll_safepoint 0
>>>> JavaThread state: _thread_blocked
>>>> at java.lang.Thread.sleep(Native Method)
>>>> at Sleeper.main(Sleeper.java:5)
>>>>
>>>> Short version: I agree that there are two "JavaThread
>>>> state: _thread_blocked"
>>>> and we should fix that. I have no preference on whether the
>>>> line is before or
>>>> after the "Thread: ..." line. Others may have an opinion.
>>>>
>>>> src/hotspot/share/runtime/vframe.cpp
>>>> I agree with the comment rewrite on L174-6.
>>> I actually kept thinking about the comment I added in lines
>>> L174-L176, but now I think it's not completely true. Because
>>> although that window of times exists there is no safepoint check
>>> inside it so the VMthread would never be able to stop the JavaThread
>>> during that time. Maybe I should remove that comment?
>>>
>>>> I like this new code:
>>>>
>>>> L182:
>>>> if(java_lang_Thread::get_thread_status(thread()->threadObj()) ==
>>>> java_lang_Thread::BLOCKED_ON_MONITOR_ENTER){
>>>> L183: wait_state = "waiting to re-lock in wait()";
>>>> L184: }
>>>> Needs space between 'if' and '(' and between ')' and '{'.
>>> Done!
>>>>
>>>> I agree with deleting old L232->250. That's the code that was
>>>> causing the confusing "waiting to re-lock in wait()" output
>>>> in the frame(s) after frame 0.
>>>>
>>>> Thumbs up!
>>>>
>>>> As David says, having some example current and new output in the
>>>> bug report
>>>> would help.
>>> Done!
>>>
>>>
>>> Thanks!
>>> Patricio
>>>
>>>> Dan
>>>>
>>>>
>>>>> Bug URL: https://bugs.openjdk.java.net/browse/JDK-8150689
>>>>>
>>>>> Thanks,
>>>>> Patricio
>>>>
>>>
>>>
>>
>
More information about the hotspot-runtime-dev
mailing list