[OpenJDK 2D-Dev] <AWT Dev> RFR: 8240654 : ZGC can cause severe UI application repaint issues
philip.race at oracle.com
Thu Jun 11 15:12:54 UTC 2020
Blitting a full screen window for me (slightly under 1920x1080) is a
little under 3 ms.
With no need to copy / retry it is a just around 1 ms which I am not
sure whether to trust completely.
On another machine with a finer grained timer it was > 1 ms.
On 6/10/2020 6:50 PM, Sergey Bylokhov wrote:
> Just curious about what performance degradation we will have in the
> worst case.
> On 6/10/20 3:57 pm, Kevin Rushforth wrote:
>> Code changes look good.
>> I verified that the bug happens on my Windows 10 laptop with
>> SwingSet2 and with the LargeWindowPaintTest without the patch, and
>> everything looks good with the patch.
>> -- Kevin
>> On 6/10/2020 1:48 PM, Philip Race wrote:
>>> Bug: https://bugs.openjdk.java.net/browse/JDK-8240654
>>> Webrev: http://cr.openjdk.java.net/~prr/8240654/index.html
>>> This is for JDK 15 so review ASAP please since RDP 1 and the test
>>> cycle are looming.
>>> This is not a fix for a JDK bug. It is a bunch of workarounds for a
>>> Microsoft Windows bug affecting
>>> GDI in the context of ZGC (http://openjdk.java.net/jeps/333).
>>> Some extra details about the Windows bug at the end, but first the
>>> technical details of the fix.
>>> With ZGC's memory allocation requirement of reserving memory in 2Mb
>>> chunks some Windows GDI
>>> functions, mostly involving some bitmaps APIs may return a failure
>>> code (ie fail!)
>>> This typically occurs when Java heap memory is used for a Java image
>>> and then in a JNI
>>> call we use GetPrimitiveArrayCritical so that Java heap allocated
>>> memory is passed to a GDI
>>> function AND the Java heap memory spans one of the 2Mb boundaries.
>>> This is very easy to trigger in almost any Java UI app if the window
>>> is of a large enough (ie typical) size.
>>> NB: if you have an Nvidia or ATI card, then you won't see it,
>>> because the D3D pipeline doesn't
>>> call the affected method but if you have an Intel chip as do 90% (?)
>>> of laptops you will see it.
>>> There are also several other places we found that are affected.
>>> Printing is the other one
>>> somewhat easy to trigger. The others : custom cursors and tray icons
>>> are less common.
>>> The painful thing here is that there is no definitive list (a list
>>> of the known ones is below) of
>>> affected Windows GDI APIs and we are just hunting around our code
>>> trying to see where it
>>> might be side-swiped by this bug.
>>> The basic approach in these workarounds is that for cases where
>>> performance does not matter we now copy
>>> and for cases where performance does matter or larger amounts of
>>> memory is involved we check if
>>> the return value of the GDI function indicates failure and then
>>> re-try with a copy of the heap memory.
>>> Unless GDI was randomly failing already (unlikely) this should be a
>>> no-risk solution in the high profile cases.
>>> We have done performance measurements on the important screen case
>>> and the failures
>>> happen fast so the penalty is then in the re-try which is only if
>>> ZGC is enabled.
>>> Always copying the memory is slower (and memcpy is the slow
>>> operation) than an alternative approach
>>> that "knows" about the memory allocation of ZGC but this coupling
>>> and the complexity seem like they aren't
>>> worth it since I haven't seen any visible performance consequence.
>>> That can be revisited
>>> some day if need be, but for now we have correctness which is the
>>> key as well as sufficient performance.
>>> I've created an automated test for the most important on-screen case.
>>> Also a manual printing test case which invokes ZGC is provided since
>>> there we also only
>>> conditionally copy. In the other cases we now always copy so
>>> existing test cases should over those.
>>> There is some clean up in this fix - one completely unused (provably
>>> so because it was #if'd out)
>>> JNI method in awt_PrintJob.cpp is removed since it had code that
>>> looked like it needed a workaround,
>>> which would be somewhat of a waste of effort.
>>> the doPrintBand code and its callee bitsToDevice has code I think we
>>> can remove too since
>>> I don't see how it ever gets executed (the top down case for
>>> browserPrint == true) but
>>> I think I'll save that for a P4 follow-on since it does nothing that
>>> would be affected by this
>>> Windows bug.
>>> One oddity is the in the printing case I observed that some times
>>> the rendering is performed
>>> even if an error code is returned. I don't know why, but in code we
>>> can't tell that it was actually
>>> rendered and in any case there is no harm in repeating the call with
>>> copied memory.
>>> We are right before the JDK15 stabilisation fork and this fix needs
>>> to go there and will
>>> but the webrev is against jdk/client simply because jdk15 does not
>>> exist yet !
>>> Please test and review ASAP.
>>> About the bug:
>>> Microsoft has acknowleged the bug and will publish a knowledge base
>>> article about it
>>> but a fix may show up only in a future version of Windows. Not, it
>>> seems, any time soon.
>>> Below is a list of potentially affected GDI APIs. Per microsoft
>>> whether it actually manifests in
>>> any specific case depends on "branching"
More information about the 2d-dev