RFR(S): 8236938: [TESTBUG] JFR event MetaspaceAllocationFailure is not tested

[Thomas Stüfe]

Hi Mikhailo,

On Sat, Feb 29, 2020 at 1:05 AM <mikhailo.seledtsov at oracle.com> wrote:

> Hi Thomas,
>
>    Thank you for review. Please see my comments inline.
> On 2/27/20 11:15 AM, Thomas Stüfe wrote:
>
> Hi Mikhailo,
>
> looks good to me. Cannot comment on the effects of the class renaming.
>
> The test is of course vulnerable against changes in the metaspace usage of
> the base jdk but I do not see how that could be changed. Make sure you test
> with and without CDS enabled since if CDS is not enabled or does not work
> the base JDK will use a lot more Metaspace and 20M may not be enough. Or,
> make sure CDS is always enabled and working.
>
> Good point, thanks.
>
> I have added -Xshare:off to the @run params to make the test more
> deterministic. And I quickly ran into OOME:Metaspace, hence I increased the
> MaxMetaspaceSize to 100M.
>
> Here is a new @run statement:
>
>  * @run main/othervm -Xmx1G -XX:MaxMetaspaceSize=100M
>  *      -XX:StartFlightRecording -Xshare:off
>  *      jdk.jfr.event.runtime.TestMetaspaceAllocationFailure
>  */
>
>
> That is actually better than what I proposed.


> If you want to be really thorough you may want to check exhaustion on
> -XX:CompressedClassSpaceSize (with infinite MaxMetaspaceSize) too. As a
> rough guide, each loaded class takes between 512 and 1K of space.
>
>  Thanks. I have added another @run to the same test using
> CompressedClassSpaceSize:
>
>  * @run main/othervm -Xmx1G -XX:CompressedClassSpaceSize=100M
>  *      -XX:StartFlightRecording -Xshare:off
>  *      jdk.jfr.event.runtime.TestMetaspaceAllocationFailure
>
> It works, I am able to provoke the MetaspaceAllocationFailure JFR event.
>
> I will do some more testing on multiple platforms.
>
>
> Great.

> Thank you,
>
> Misha
>
>
> ..Thomas



> Cheers, Thomas
>
> On Thu, Feb 27, 2020 at 4:51 PM <mikhailo.seledtsov at oracle.com> wrote:
>
>> Please review this new test for JFR event MetaspaceAllocationFailure.
>>
>> The test fills the metaspace by generating classes in a loop, until the
>> desired event is delivered to the receiver.
>> By creating a new instance of class loader for each iteration test
>> allows metaspace to be cleaned up once MetaspaceAllocationFailure
>> condition is detected, thus not causing full-blown OOME. However, on
>> occasion, it may happen, hence the try/catch clause.
>> Ran this at least 100 times, works well and reliably.
>>
>> Also moved the class generator utility from Runtime test library to
>> common test library; renamed it to avoid name clash with
>> another already existing GeneratingClassLoader.
>>
>>      JBS: https://bugs.openjdk.java.net/browse/JDK-8236938
>>      Webrev: http://cr.openjdk.java.net/~mseledtsov/8236938.01/
>>      Testing:
>>          1. Ran over 100 times on Linux, Windows and MAC: All PASS
>>          2. Ran all runtime tests that rely on the class loader
>> generator runtime library.
>>
>>
>> Thank you,
>> Misha
>>
>>