RFR: 8365588: defineClass that accepts a ByteBuffer does not work as expected [v2]
Xueming Shen
sherman at openjdk.org
Tue Sep 30 21:02:03 UTC 2025
> ### Background
>
> - ClassLoader.defineClass can receive class data in the form of arrays or ByteBuffers.
> - For array-backed data (defineClass1), a defensive copy is made before passing it to JVM_DefineClassWithSource().
> - For Direct-ByteBuffer variants (defineClass2), no defensive copy is made, which creates a risk that the underlying bytes could be modified while the JVM is processing them.
> - Although a caller could always modify a buffer before a defensive copy is made — a race condition that cannot be completely prevented — the **_main concern_** is ensuring that the JVM never processes class bytes that are being concurrently modified.
>
> ### Problem
>
> - Concurrent modification risk during processing: while we cannot prevent pre-copy modifications, we **_must prevent the JVM from using class bytes that are being modified concurrently._**
> - Performance concerns: defensive copies have a cost, especially for direct byte buffers. Making copies unnecessarily for trusted class loaders (like the built-in class loader) would hurt performance.
>
> ### Solution
>
> - Make a defensive copy of the direct byte-buffer only when the class loader is **NOT** a built-in/trusted class loader.
> - For the built-in class loader, skip the copy because the JVM can guarantee that the buffer contents remain intact.
>
> This approach ensures the integrity of class bytes processes for untrusted or custom class loaders while minimizing performance impact for trusted or built-in loaders.
>
> ### Benchmark
>
> A JMH benchmark has been added to measure the potential cost of the defensive copy. The results indicate that the performance impact is minimal and largely insignificant.
>
> **Before:**
>
>
> Benchmark Mode Cnt Score Error Units
> ClassLoaderDefineClass.testDefineClassByteBufferDirect avgt 15 8387.247 ± 1405.681 ns/op
> ClassLoaderDefineClass.testDefineClassByteBufferHeap avgt 15 8971.739 ± 1020.288 ns/op
> Finished running test 'micro:org.openjdk.bench.java.lang.ClassLoaderDefineClass'
> Test report is stored in /Users/xuemingshen/jdk26/build/macosx-aarch64/test-results/micro_org_openjdk_bench_java_lang_ClassLoaderDefineClass
>
>
> **After:**
>
>
> Benchmark Mode Cnt Score Error Units
> ClassLoaderDefineClass.testDefineClassByteBufferDirect avgt 15 8521.881 ± 2002.011 ns/op
> ClassLoaderDefineClass.testDefineClassByteBufferHeap avgt 15 9002.842 ± 1099.635 ns/op
> Finished running test 'mi...
Xueming Shen has refreshed the contents of this pull request, and previous commits have been removed. The incremental views will show differences compared to the previous content of the PR. The pull request contains one new commit since the last revision:
8365588: defineClass that accepts a ByteBuffer does not work as expected
-------------
Changes:
- all: https://git.openjdk.org/jdk/pull/27569/files
- new: https://git.openjdk.org/jdk/pull/27569/files/cb294829..c7cb474f
Webrevs:
- full: https://webrevs.openjdk.org/?repo=jdk&pr=27569&range=01
- incr: https://webrevs.openjdk.org/?repo=jdk&pr=27569&range=00-01
Stats: 7 lines in 2 files changed: 0 ins; 0 del; 7 mod
Patch: https://git.openjdk.org/jdk/pull/27569.diff
Fetch: git fetch https://git.openjdk.org/jdk.git pull/27569/head:pull/27569
PR: https://git.openjdk.org/jdk/pull/27569
More information about the core-libs-dev
mailing list