[jdk17] RFR: 8269897: Shenandoah: Treat UNKNOWN refs access as strong
shade at openjdk.java.net
Wed Jul 7 10:44:51 UTC 2021
On Tue, 6 Jul 2021 19:11:58 GMT, Roman Kennke <rkennke at openjdk.org> wrote:
> We've observed test failures in jcstress, see:
> We used to treat UNKNOWN reference accesses like weak accesses. UNKNOWN is used for Unsafe, reflection and JNI accesses, where it cannot be determined at compilation-time if we are accessing a regular field or a Reflection.referent field. The rationale for treating UNKNOWN as weak was that if the reference is a regular reference, then the value would be strongly reachable anyway, and only if it is a referent field would reachability matter. However, it turns out that this assumption is wrong: the test shows that a reference that is only weakly reachable can be legitimately written into a field, thus resurrecting the reference, and when that weakly reachable reference is loaded, it would be (wrongly) filtered as NULL.
> A fix is to treat UNKNOWN accesses as strong. Accessing Reference.referent via reflection, JNI or Unsafe is Bad Idea anyway.
> This test shows the problem with CAS, but I believe it affects all accesses via reflection, JNI, etc.
> - [x] the provided jcstress test
> - [x] hotspot_gc_shenandoah
> - [x] tier1
It looks okay to me, and it fixes the jcstress failures. But I see one of the `tier1` tests failing:
$ make run-test TEST=gc/TestReferenceClearDuringReferenceProcessing.java TEST_VM_OPTS="-XX:+UseShenandoahGC"
TEST RESULT: Failed. Execution failed: `main' threw exception: java.lang.RuntimeException: ref not enqueued
I wonder if that failure indicates that fix is still incomplete or too conservative.
More information about the hotspot-gc-dev