<AWT Dev> <Awt Dev> [9] Review Request for 8132664: closed/javax/swing/DataTransfer/DefaultNoDrop/DefaultNoDrop.java locks on Windows

[Sergey Bylokhov]

On 03.09.15 17:43, Semyon Sadetsky wrote:
>
>
> On 8/5/2015 2:33 PM, Sergey Bylokhov wrote:
>> On 05.08.15 14:20, Semyon Sadetsky wrote:
>>>
>>>
>>> On 8/5/2015 1:39 PM, Sergey Bylokhov wrote:
>>>> On 05.08.15 13:18, Semyon Sadetsky wrote:
>>>>>
>>>>>
>>>>> On 8/5/2015 12:27 PM, Sergey Bylokhov wrote:
>>>>>> On 04.08.15 14:54, Semyon Sadetsky wrote:
>>>>>>> On 8/3/2015 6:05 PM, Sergey Bylokhov wrote:
>>>>>>>> Hi, Semyon
>>>>>>>> Did you try to change dwMilliseconds from INFINITE to the
>>>>>>>> timeout(10 seconds by default?) which is passed to the method?
>>>>>>>> It does not help? Because even when dnd is not used we should
>>>>>>>> not wait event for infinite time.
>>>>>>> It would not help to fix the issue because 10 seconds is too big
>>>>>>> interval. But for consistency it is not bad to have a time limit.
>>>>>>> http://cr.openjdk.java.net/~ssadetsky/8132664/webrev.01/
>>>>>> Note that syncNativeQueue is intended to wait until the native
>>>>>> event queue is flushed or until timeout will expire. So even if
>>>>>> timeout is expired we collect the native events during this period
>>>>>> of time.Can you double check that the event counter is incremented
>>>>>> during dnd? I do not know how we block the toolkit thread,
>>>>>> probably we create some nested loops which ignore our event posted
>>>>>> from syncNativeQueue, can we change that?
>>>>> Yes, this is an internal secondary loop which waits for mouse
>>>>> release event.
>>>> Can we change the condition and process the sync event in this loop?
>>> Why? Will receive all events on the toolkit thread when doDragDrop
>>> returns.
>>
>> When how we get dragenter/exit events?
>
> On the platform side they are not events but callbacks which are
> converted into events on the java side and added to the queue. So they
> will be detected by isEQEmpty() and waitForIdle() will be repeated.

This is not directly related to this fix. Most of our callbacks/events 
posts events to the EQ, but there is a possibility for a lag between 
callback and a post events, and this is why the syncNativeQueue was added.
Let's return to the beginning: the syncNativeQueue method according to 
its specifications should try to flush the native system, track the 
native activity and should not wait more than timeout parameter.

>
>>
>>>>> Event counter is not changed during toolkit thread blocking of
>>>>> cause. Not sure that we can change that. But since toolkit queue is
>>>>> blocked we can assume that we are synced.
>>>>>>
>>>>>> The timeout value is maintained on the shared level and actually
>>>>>> this test will fail with timeout on osx as well JDK-7185258. The
>>>>>> test will fail even if the time out will be changed to ±100ms,
>>>>>> because it call realsync on each pixel move, ±200 times. This
>>>>>> problem can be fixed in different ways like tweak of timeouts and
>>>>>> numbers of iterations, or changing the test to call w4idle only
>>>>>> after the latest move(actually I think this method can be moved to
>>>>>> the robot class).
>>>>>>
>>>>>> So I still think that the right fix for a deadlock, which is
>>>>>> subject of this bug, is simply change the syncNativeQueue and
>>>>>> waits using a timeout if it is positive, and waits forever if
>>>>>> timeout is negative (the same bug on osx JDK-8080504).
>>>>> I'm not sure that waiting brings any value. What do you propose to
>>>>> return if it timed out? The event counter will not be changed
>>>>> regardless of waiting.
>>>> But it should be changed, because we get native events from the
>>>> system during dnd and in each such callback we should update this
>>>> counter. If callbacks were not called=>counter was not updated then
>>>> sync assume that currently we do not process events. If callbacks
>>>> were called then sync assume that we have events in the native queue
>>>> and should try to sync again on the next iteration.
>>> No. Events are not processed while toolkit is blocked in
>>> doDragDrop(). The application state is frozen on that period. That is
>>> Windows approach.
>>>>
>>>>> With such waiting the test will fail because of either jtreg timeout
>>>> Default timeout is 120 seconds for everything, the test try to sync
>>>> the queue 200 of times after each move, so yes it can fail with
>>>> timeout even if spend in nativesync 200 ms, the possible solutions
>>>> were in my previous email.
>>>>> either InfiniteLoop exception.
>>>> This exception will be disabled by default lately in jdk9 timeframe,
>>>> right now it helps to find some related issues.
>>>>>
>>>>>>>
>>>>>>>>
>>>>>>>> On 03.08.15 17:26, Semyon Sadetsky wrote:
>>>>>>>>> Hello,
>>>>>>>>>
>>>>>>>>> Please review fix for JDK9:
>>>>>>>>>
>>>>>>>>> bug: https://bugs.openjdk.java.net/browse/JDK-8132664
>>>>>>>>> webrev: http://cr.openjdk.java.net/~ssadetsky/8132664/webrev.00/
>>>>>>>>>
>>>>>>>>> DoDragDrop() is blocking, so upon drag operation is triggered
>>>>>>>>> the toolkit thread is blocked and the WM_AWT_WAIT cannot be
>>>>>>>>> processed which in its turn blocks the AWT robot.
>>>>>>>>> The solution is to escape AWT robot waiting in
>>>>>>>>> syncNativeQueue() if drag operation is in progress.
>>>>>>>>>
>>>>>>>>> --Semyon
>>>>>>>>
>>>>>>>>
>>>>>>>
>>>>>>
>>>>>>
>>>>>
>>>>
>>>>
>>>
>>
>>
>


-- 
Best regards, Sergey.