On Mon, 20 Jan 2025 16:48:49 GMT, Aleksey Shipilev <sh...@openjdk.org> wrote:
>> DirectByteBuffers are still using old `jdk.internal.ref.Cleaner` >> implementation. That implementation carries a doubly-linked list, and so >> makes DBB suffer from the same issue fixed for generic >> `java.lang.ref.Cleaner` users with >> [JDK-8343704](https://bugs.openjdk.org/browse/JDK-8343704). See the bug for >> the reproducer. >> >> We can migrate DBBs to use `java.lang.ref.Cleaner`. >> >> There are two pecularities during this rewrite. >> >> First, the old ad-hoc `Cleaner` implementation used to exit the VM when >> cleaning action failed. I presume it was to avoid memory leak / accidental >> reuse of the buffer. I moved the relevant block to `Deallocator` directly. >> Unfortunately, I cannot easily test it. >> >> Second is quite a bit hairy. Old DBB cleaning code was hooked straight into >> `Reference` processing loop. This was possible because we could infer that >> the weak references we are processing were DBB cleaning actions, since old >> `Cleaner` was the only use of this code. With standard `Cleaner`, we have >> lost this association, and so we cannot really do this from the reference >> processing loop. With the patched version, we now rely on normal `Cleaner` >> thread to do cleanups for us. Because of this, there is a new outpacing >> opportunity window where reference processing might have been over, but >> cleaner thread has not reacted yet. >> >> Tests show that DirectBufferAlloc tests are still surviving this, possibly >> due to exponential sleep-backoff already built in. See the reclamation path >> in `Bits.unreserveMemory`: >> https://github.com/openjdk/jdk/blob/c207cc7e705d3f449f2387324d86cfb31ce40c44/src/java.base/share/classes/java/nio/Bits.java#L106-L186 >> >> Additional testing: >> - [x] Linux x86_64 server fastdebug, `java/nio java/io` >> - [x] Linux AArch64 server fastdebug, `java/nio java/io` >> - [ ] Linux x86_64 server fastdebug, `all` >> - [ ] Linux AArch64 server fastdebug, `all` > > Aleksey Shipilev has updated the pull request incrementally with one > additional commit since the last revision: > > Visibility and whitespace Marked as reviewed by uschindler (Author). This looks fine for me. Older Lucene versions won't break, as they use `sun.misc.Unsafe#invokeCleaner()` which behaves as before. Code that deep reflected and made `DirectByteBuffer#cleaner()` accessible was broken for long time anyways (since Java 9), so the new names and class types should not affect uptodate code. I'd remove the `catch (Throwable)` => exit VM code completely. The called methods have no checked exceptions. If the deallocation fails for some (internal) reason it its already broken and a crush will come sooner or later anyways. ------------- PR Review: https://git.openjdk.org/jdk/pull/22165#pullrequestreview-2562957601 PR Comment: https://git.openjdk.org/jdk/pull/22165#issuecomment-2603003680
