Re: Review Request For 8146230: Crash in JNI_ArgumentPusherVaArg::JNI_ArgumentPusherVaArg(_jmethodID*, __va_list_tag*)+0xa

[Semyon Sadetsky]

Hi Ambarish,

 Thank you for removing the lock.
But I still do not understand the root cause. waitForEvents() is only 
called on the toolkit thread How it could be executed concurrently?

Are you able to reproduce the issue? Are you sure that correcting method 
IDs initialization really fixes it?

--Semyon


On 7/8/2016 4:24 PM, Ambarish Rapte wrote:

Hi Semyon,

                Thanks for the review.

                Please review the webrev.01, updated as per the review 
comments.

http://cr.openjdk.java.net/~arapte/8146230/webrev.01/ 
<http://cr.openjdk.java.net/%7Earapte/8146230/webrev.01/>

Changes:

                Initializing the thread class references in 
Java_sun_awt_X11_XToolkit_initIDs()

                As you guided,

1.LOCK on toolkit thread will be avoided by this change.

2.awtJNI_ThreadYield() is called for even when just a Frame object is 
created, which is a case of almost all basic programs.
Hence it can be early initialized.

                But regarding, XToolkit.c::get_xawt_root_shell() function.

                Verified with few sample programs with Frame, this 
function is not always called.

                Hence it can be kept as delayed initialization, when 
required.

Regards,

Ambarish

*From:*Semyon Sadetsky
*Sent:* Thursday, July 07, 2016 7:27 PM
*To:* Ambarish Rapte; Sergey Bylokhov; Alexander Scherbatiy; Prasanta 
Sadhukhan; awt-dev@openjdk.java.net
*Subject:* Re: Review Request For 8146230: Crash in 
JNI_ArgumentPusherVaArg::JNI_ArgumentPusherVaArg(_jmethodID*, 
__va_list_tag*)+0xa

On 07.07.2016 16:14, Ambarish Rapte wrote:

    Hi Semyon,

    1. The _initIDs() functions are called from static initialization
    block of particular class, hence these functions would be called
    when class is getting loaded.

         The_initIDs() functions are used to initialize the IDs & get
    value of Java class members only.

                                    For example:

    Java_sun_awt_X11_XToolkit_initIDs() initializes IDs and gets value
    from java side XToolkit class (XToolkit.java)

awtJNI_ThreadYield() is only used by the XToolkit. I don't see 
anything bad if it will be initialized there.

                2.  But awtJNI_ThreadYield()  is late initialization 
of IDs. IDs are initialized when first thread has to be yield.

Is lazy initialization really necessary? It looks like the method ID 
is used always.

ð Addition of operations to Java_sun_awt_X11_XToolkit_initIDs() is 
addition to class load time.

Lock also introduces delay. Also it maybe a source for deadlocks. 
Locking toolkit thread using global lock object should be really 
justified.

ð XToolkit.c::get_xawt_root_shell() is another example of late 
initialization of jclass, jmethodID.

Should it be fixed as well?

--Semyon

Regards,

Ambarish

*From:*Semyon Sadetsky
*Sent:* Wednesday, July 06, 2016 11:39 PM
*To:* Ambarish Rapte; Sergey Bylokhov; Alexander Scherbatiy; Prasanta 
Sadhukhan; awt-dev@openjdk.java.net <mailto:awt-dev@openjdk.java.net>
*Subject:* Re: Review Request For 8146230: Crash in 
JNI_ArgumentPusherVaArg::JNI_ArgumentPusherVaArg(_jmethodID*, 
__va_list_tag*)+0xa

Hi Ambarish,

Why not simply initialize yieldMethodID separately? For example in 
Java_sun_awt_X11_XToolkit_initIDs.

--Semyon

On 06.07.2016 14:23, Ambarish Rapte wrote:

    Hi,

                    Please review the fix for JDK9,

                    Bug: https://bugs.openjdk.java.net/browse/JDK-8146230

                    Webrev:
    http://cr.openjdk.java.net/~arapte/8146230/webrev.00/
    <http://cr.openjdk.java.net/%7Earapte/8146230/webrev.00/>

    Issue:

    1.Null pointer exception in JNI

    Cause:

    The code block was not multi thread safe.

    Issue occurs in  multi threaded , multi processor environment.

    Fix:

    1.Changed the variable used for double checking, to use
    /yieldMethodID ./

    2.Changed /yieldMethodID  to  volatile./

    3.Added AWT_LOCK() over initialize block of code.

    4.Removed unrequired  err variable.

    A drawback of Double Check Locking ( DCL ) is, if the resource
    assignment is not an atomic operation, The DCL may fail.

    But here in the code of concern, is an atomic operation. Hence DCL
    should work fine.

    Please check below reference link for more detailed discussion of
    DCL with C++.

    Verification:

    1.Tested Event tests which pass without any regression of this change.

    2.As this change only corrects existing logic, there should be no
    side effects.

    Reference:

                    C++ and the Perils of Double-Checked Locking:
    http://www.aristeia.com/Papers/DDJ_Jul_Aug_2004_revised.pdf

    Regards,

    Ambarish