RFR (XS) 8059924: com/sun/management/DiagnosticCommandMBean/DcmdMBeanPermissionsTest.java: assert(Universe::verify_in_progress() || !SafepointSynchronize::is_at_safepoint()) failed: invariant

[Aleksey Shipilev]

Hi,

There is a regression introduced by:
 https://bugs.openjdk.java.net/browse/JDK-8059595

The regression is tracked here:
 https://bugs.openjdk.java.net/browse/JDK-8059924

...and here is a fix:
 http://cr.openjdk.java.net/~shade/8059924/webrev.00/

Testing: JPRT, jdk/test/com/sun/management

The comment there should be self-explanatory. The failure requires a
precise sequence of events to happen: we should be at safepoint, and
call klass->external_name() on a previously untouched class. Before
JDK-8059595, Verifier always did klass->external_name() for all verified
classes, causing a performance trouble.

This might explain why the failure happens on com/sun/management tests
that do VM operations (GC_HeapInspection in reported failure case). This
reproduces perfectly with DcmdMBeanPermissionsTest, and the test is
fixed with the VM change above.

The alternative would be to backout the JDK-8059595 change completely,
but this gets us back to weird piggybacking on klass->external_name()
during the verification. Other alternatives: returning 0 from hashcode
during safepoint is obviously unsafe; returning zero-hashcode string
from klass->external_name() is potentially unsafe if someone is using
the name as the map key downstream.

-Aleksey.