RFR: 8261154: Memory leak in Java_java_lang_ClassLoader_defineClass0 with long class names
Claes Redestad
redestad at openjdk.java.net
Thu Feb 4 15:55:41 UTC 2021
On Thu, 4 Feb 2021 15:48:16 GMT, Thomas Stuefe <stuefe at openjdk.org> wrote:
>> This patch resolves a potential memory leak in Java_java_lang_ClassLoader_defineClass0
>>
>> I've not figured a good way to write a regression test. A crude way to manually verify the leak and the fix is provided by the added microbenchmark that triggers the malloc in ClassLoader.c and the associated leak.
>>
>> E.g., running this with `/usr/bin/time -v $BUILD_DIR/images/jdk/bin/java -Xmx256m -jar $BUILD_DIR/images/test/micro/benchmarks.jar LookupDef.*WeakClass.loadLong -f 0 -i N | grep "Maximum resident set"` yields:
>>
>> Baseline:
>> N = 20 Maximum resident set size (kbytes): 544860
>> N = 50 Maximum resident set size (kbytes): 818532
>> N = 100 Maximum resident set size (kbytes): 1388560
>> Patch:
>> N = 20 Maximum resident set size (kbytes): 480476
>> N = 50 Maximum resident set size (kbytes): 764040
>> N = 100 Maximum resident set size (kbytes): 782920
>
> Hi Cleas,
>
> looks good but why changing the the NULL comparisons?
>
> If you wanted to simplify you could completely omit the NULL checks since free(NULL) is a noop. And potentially merge the labels - where there are two - into one.
>
> Cheers, Thomas
I'm not sure every platform have always agreed free(NULL) is a noop. I suspect all the currently supported ones do, though?
-------------
PR: https://git.openjdk.java.net/jdk/pull/2407
More information about the core-libs-dev
mailing list