RFR(M): 8069412 : Locks need better debug-printing support

David Chase david.r.chase at oracle.com
Tue Feb 17 13:22:10 UTC 2015 

This time for sure...

http://cr.openjdk.java.net/~drchase/8069412/webrev.04/

Fixed style notes from Vladimir and David,
reworked log_lock_optimization to take a Compile *, modified callers,
therefore also reworked the debugging-general form of is_nested_lock_region to take a Compile *
and its callers.

Retested w/ jtreg -XX:+LogCompilation to be sure I had not typoed,
also reran one of my one-off tests for lock optimization and checked its logged+formatted output.

David

On 2015-02-16, at 7:15 PM, Vladimir Kozlov <vladimir.kozlov at oracle.com> wrote:

> Yes, Compile is Phase's subclass but 'C' field is NULL:
> 
> Phase::Phase( PhaseNumber pnum ) : _pnum(pnum), C( pnum == Compiler ? NULL : Compile::current()) {
> 
> So you can't pass C as phase into log_lock_optimization() and is_nested_lock_region(). I suggest to use Compile*.
> 
> Style is not fixed:
> 
> +    if (monitor->eliminated())
> +        if(is_compiled_frame())
> +            tty->print(" ( lock is eliminated in compiled frame )");
> +        else
> +            tty->print(" ( lock is eliminated, frame not compiled )");
> 
> Vladimir
> 
> On 2/16/15 10:09 AM, David Chase wrote:
>> Improved webrev, should address issues:
>> 
>> http://cr.openjdk.java.net/~drchase/8069412/webrev.03/
>> 
>> I did some more work on LogCompilation -- testing (there is none for that tool in jtreg or JPRT)
>> revealed that the 9+changes tool would crash on 8u logs, and that (1) made it hard for me to test
>> my changes against existing logs known to contain new-features data and (2) seems kinda graceless
>> anyhow.
>> 
>> David
>> 
>> On 2015-02-13, at 11:50 AM, Vladimir Kozlov <vladimir.kozlov at oracle.com> wrote:
>> 
>>> On 2/13/15 5:33 AM, David Chase wrote:
>>>>> In general I am fine with this changes. There are styles issues, indention (used 4 bytes instead of 2). Other then that I have only few comments.
>>>> 
>>>> markOop.cpp?  Fixed it.
>>> 
>>> Also basicLock.cpp and vframe.cpp. They also miss {} for new if statements.
>>> 
>>>> LogCompilation files are java and 4-character indent by default.
>>>> 
>>>>> In vframe.cpp why you cloned print_locked_object_class_name() to all passes. Instead you could set Boolean local in inner scope if you want to print Verbose output after print_locked_object_class_name().
>>>> 
>>>> "mark" was not in scope outside, it seemed cleaner on balance just to clone the call.
>>> 
>>> You can set 'markOop mark = NULL' before 'if (!found_first_monitor' and use 'if (Verbose && (mark != NULL))' condition to print additional info. I don't think to have 3 copy of code is clean.
>>> 
>>>> 
>>>>> Also I don't think you need "_debug" in then name LockNode::is_nested_lock_region_debug().
>>>> 
>>>> I wanted it to be clear that this was only for debugging -- it is clone of the previous method with tracing in the failure paths.
>>> 
>>> Why duplicate code that to do logging. Note, the logging is available in product VM too. I would suggest to modify the original method to do logging. You can pass additional bool argument to trigger logging if you don't want to do it always.
>>> 
>>>> 
>>>>> You only need to add additional check is_Lock() in debug mode where before we may only check jvms() != NULL. I think you simply can add explicit check in macro.cpp code:
>>>>> 
>>>>> !       // Not that preceding claim is documented anywhere else.
>>>>> !       if (alock->jvms() != NULL && alock->is_Lock()) {
>>>>> !         if (alock->as_Lock()->is_nested_lock_region()) {
>>>> 
>>>> I didn't change jvms() assignment -- for Unlock nodes there is a debug-only second field to hold this information,
>>>> because this is not the only place sensitive to which AbstractLock nodes have jvms() attached, and adding a jvms()
>>>> to Unlock nodes caused additional failures.  So the original condition still works.
>>> 
>>> Okay, I got it.
>>> 
>>> Thanks,
>>> Vladimir
>>> 
>>>> 
>>>> thanks for the review,
>>>> 
>>>> David
>>>> 
>>>>> Thanks,
>>>>> Vladimir
>>>>> 
>>>>> On 2/12/15 7:09 PM, David Chase wrote:
>>>>>> http://cr.openjdk.java.net/~drchase/8069412/webrev.00/
>>>>>> 
>>>>>> This is the debugging support that I found very helpful to make progress on
>>>>>> https://bugs.openjdk.java.net/browse/JDK-8066576 "Lock still held"
>>>>>> 
>>>>>> It comes in three parts:
>>>>>> 
>>>>>> 1) When -XX:+Verbose, locks in stack backtrace are printed with additional information,
>>>>>> like so:
>>>>>> 
>>>>>> "FooThread" #19 daemon prio=10 os_prio=31 tid=0x00007fe41c2c8000 nid=0x6203 waiting for monitor entry [0x0000000110072000]
>>>>>>    java.lang.Thread.State: BLOCKED (on object monitor)
>>>>>>    JavaThread state: _thread_blocked
>>>>>> Thread: 0x00007fe41c2c8000  [0x6203] State: _at_safepoint _has_called_back 0 _at_poll_safepoint 0
>>>>>>    JavaThread state: _thread_blocked
>>>>>>    at foo.Foo.send(Foo.java:21)
>>>>>>    - waiting to lock <0x0000000780b89008> (a java.lang.Object)
>>>>>>      lockbits= locked(0x00007fe41c23d09a)->monitor={count=0x0000000000000001,waiters=0x0000000000000000,recursions=0x0000000000000000,owner=0x00007fe41b8fc000}
>>>>>>    at foo.Bar.bar(Bar.java:1)
>>>>>> 
>>>>>> The new line is the one beginning "     lockbits".
>>>>>> Without -XX:+Verbose, the old behavior remains.
>>>>>> 
>>>>>> 2) In a debug (#ifdef ASSERT) build, when -XX+:LogCompilation is specified,
>>>>>> additional records are written to the compilation log file describing which lock optimizations
>>>>>> are performed (and in some cases, which are not performed, and why).
>>>>>> 
>>>>>> For example:
>>>>>> 
>>>>>>   <eliminate_lock_set_nested lock='0' compile_id='1840' class_id='unlock' kind='?' stamp='350.736'>
>>>>>>   </eliminate_lock_set_nested>
>>>>>>   <eliminate_lock_set_nested lock='1' compile_id='1840' class_id='lock' kind='coarsened' stamp='350.736'>
>>>>>>   </eliminate_lock_set_nested>
>>>>>> 
>>>>>> where the first record indicates that a previously unoptimized lock (kind='?') was set to 'nested'
>>>>>> and the second record indicates that a previously coarsened lock was set to nested.
>>>>>> Other additional information recorded includes locations for Unlock nodes.
>>>>>> 
>>>>>> Even in release builds, the kind of elimination (nested, coarsened, non-escaping) is now noted.
>>>>>> 
>>>>>> 3) The compilation processing tool was enhanced to process these new records, producing output like this:
>>>>>> 
>>>>>> 1783 eliminate_lock_NOT_nested_lock_region ? lock  98.713 [@36 foo.Foo::fooMethod (330 bytes)]
>>>>>> 1783 eliminate_lock_INLR_1 ? lock  98.713 [@36 foo.Foo::fooMethod (330 bytes)]
>>>>>> 1783 eliminate_lock_set_nested ? unlock  98.713 [@-1 bar.BAR::setData (53 bytes), @54 foo.Foo::fooMethod (330 bytes)]
>>>>>> 1783 eliminate_lock_set_nested ? lock  98.713 [@-1 bar.BAR::setData (53 bytes), @54 foo.Foo::fooMethod (330 bytes)]
>>>>>> 1783 eliminate_lock_set_nested nested lock  98.713 [@-1 bar.BAR::setData (53 bytes), @54 foo.Foo::fooMethod (330 bytes)]
>>>>>> 1783 eliminate_lock_set_nested ? unlock  98.713 [@-1 bar.BAR::getLength (5 bytes), @65 foo.Foo::fooMethod (330 bytes)]
>>>>>> 1783 eliminate_lock_set_nested ? lock  98.713 [@-1 bar.BAR::getLength (5 bytes), @65 foo.Foo::fooMethod (330 bytes)]
>>>>>> 1783 eliminate_lock_set_nested nested lock  98.713 [@-1 bar.BAR::getLength (5 bytes), @65 foo.Foo::fooMethod (330 bytes)]
>>>>>> 1783 eliminate_lock_NOT_nested_lock_region ? lock  98.713 [@202 foo.Foo::fooMethod (330 bytes)]
>>>>>> 1783 eliminate_lock_INLR_1 ? lock  98.713 [@202 foo.Foo::fooMethod (330 bytes)]
>>>>>> 1783 eliminate_lock nested unlock  98.713 [@-1 bar.BAR::getLength (5 bytes), @65 foo.Foo::fooMethod (330 bytes)]
>>>>>> 1783 eliminate_lock nested lock  98.713 [@-1 bar.BAR::getLength (5 bytes), @65 foo.Foo::fooMethod (330 bytes)]
>>>>>> 1783 eliminate_lock nested unlock  98.714 [@-1 bar.BAR::setData (53 bytes), @54 foo.Foo::fooMethod (330 bytes)]
>>>>>> 1783 eliminate_lock nested lock  98.714 [@-1 bar.BAR::setData (53 bytes), @54 foo.Foo::fooMethod (330 bytes)]
>>>>>> 
>>>>>> This was tested running jtreg on compiler and runtime,
>>>>>> as well as many runs of the problematic application for 8066576,
>>>>>> plus JPRT of the hotspot testsuite,
>>>>>> plus built with XCode 4.6.3 and 6.1.1 on Mavericks and gcc 4.8.2 on Ubuntu 14.04.
>>>>>> 
>>>>>> David
>>>>>> 
>>>>>> 
>>>> 
>> 

-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 801 bytes
Desc: Message signed with OpenPGP using GPGMail
URL: <http://mail.openjdk.java.net/pipermail/hotspot-compiler-dev/attachments/20150217/3702930e/signature.asc>

More information about the hotspot-compiler-dev mailing list