RFR: 8291897: TerminatingThreadLocal(s) not registered from virtual thread(s)

[Alan Bateman]

On Thu, 4 Aug 2022 10:57:53 GMT, Peter Levart <plevart at openjdk.org> wrote:

> This is an attempt to fix inconsistent behavior of TerminatingThreadLocal(s) when used internally in JDK for per-carrier-thread caches of native ByteBuffer(s) and NativeBuffer(s). If used from virtual thread, such TerminatingThreadLocal(s) are not registered with the carrier thread for the cleanup.
> The fix introduces an internal CarrierThreadLocal subclass of ThreadLocal which rewires its API methods to use current thread's carrier thread as the source of ThreadLocalMap instead of current thread itself (if it is a VirtualThread for example). TerminatingThreadLocal is now a subclass of CarrierThreadLocal. It seems only per-carrier-thread caching is a usecase for it. The uses of JavaLangAccess in various places to access a carrier-thread value of given ThreadLocal has been replaced by public API calls and the use of CarrierThreadLocal instead of plain ThreadLocal. JavaLangAccess is still used to dispatch the CarrierThreadLocal API methods, but only for that. Would someone be tempted to use JavaLangAccess methods directly, they now require CarrierThreadLocal argument to guard against missuses.
> The REGISTRY of TerminatingThreadLocal(s) that tracks which of them have values bound to a particular carrier thread is now implemented conveniently  with a CarrierThreadLocal.
> The test is expanded with a case that demonstrates a situation where a carrier thread is terminated. Since it must wait for 30 seconds for that to happen, only one of the test cases is performed in this mode. The correct logic of TerminatingThreadLocal is still verified with all test-cases using platform threads that can be terminated more rapidly.

I did a quick pass over this and it looks good. Adding CarrierThreadLocal is a good idea and as it allows the use-sites to be revert/cleaned up. Just a few minor comments so far.

src/java.base/share/classes/sun/nio/ch/Util.java line 233:

> 231:         }
> 232: 
> 233:         BufferCache cache = bufferCache.get();

I assume JLA is no longer needed.

test/jdk/jdk/internal/misc/TerminatingThreadLocal/TestTerminatingThreadLocal.java line 40:

> 38:  * @compile --enable-preview -source ${jdk.version} TestTerminatingThreadLocal.java
> 39:  * @run main/othervm --enable-preview -Djdk.virtualThreadScheduler.parallelism=1 -Djdk.virtualThreadScheduler.maxPoolSize=2 TestTerminatingThreadLocal
> 40:  */

If you add `@enablePreview` then it will allow you to drop the `@compile` line and drop `--enable-preview from the `@run` tag. 

You might want to add the bugId to the `@bug` list too.

-------------

PR: https://git.openjdk.org/jdk/pull/9743