Mercurial > hg > Members > kono > jpf-core
diff src/main/gov/nasa/jpf/jvm/bytecode/RETURN.java @ 24:6774e2e08d37
the fix I would have liked to avoid - apparently hotspot internally does nested locking during class init, which can lead to deadlocks such as described in http://ternarysearch.blogspot.ru/2013/07/static-initialization-deadlock.html. Actually, it's not a regular deadlock since core dumps still list the threads as runnable, althouth it doesn't seem to be a livelock either. In any case, it can be simulated by nested locking and clinit execution, and it is such a serious defect that we want to be able to catch it. The general mechanism is to replace the disparate (but properly ordered) direct clinit calls of the generic ClassInfo.initializeClass() with a single sythetic method that includes all required locking (bottom up), clinit calls / class status change (top down), and unlocking (top down). We also need to add a synthetic insn to defer changing the class status of classes that don't have clinits(), or otherwise the correct lock/unlock order will not amount to anything if the hierarchy is entered through one of the clinit-absent classes. Now we get proper deadlocks if there are concurrent cyclic dependencies during class resolution. However, this can be such a state exploder that we certainly don't want this as the default behavior, especially since it probably is hotspot specific. Nested class init locking is therefore controlled by jvm.nested_init and respective jvm.nested_init.include/exclude options. Added a NestedInitTest to demonstrate use. Thanks to Lilia Abdulina for bringing this long forgotten issue up
In the wake of nested locks, there were a number of cases to fix that implicitly relied on absent clinits because clients were not properly checking for re-execution (most notably java.util.Exchanger). This mostly came in through MJIEnv.newObject/ElementInfo. We might turn ClinitRequired into a handled exception at some point, to catch such cases during compilation.
Added a UnknownJPFClass exception (in analogy to ClinitRequired), to make clients aware of failed class load attempts/reasons.
fixed Exchanger peer, which was not giving up the lock when timing out. This is an example of a lockfree wait op that can time out. Basically, ThreadInfo.isWaiting() needs to be complemented by a isWaitingOrTimedOut(), and ElementInfo.notifies0() has to be aware of it
fixed NPE when setting report.probe_interval in tests, which was missing that it had to create a stat object
author | Peter Mehlitz <Peter.C.Mehlitz@nasa.gov> |
---|---|
date | Tue, 21 Apr 2015 00:34:15 -0700 |
parents | 61d41facf527 |
children | 8aded593a50f |
line wrap: on
line diff
--- a/src/main/gov/nasa/jpf/jvm/bytecode/RETURN.java Wed Apr 15 22:40:21 2015 -0700 +++ b/src/main/gov/nasa/jpf/jvm/bytecode/RETURN.java Tue Apr 21 00:34:15 2015 -0700 @@ -33,12 +33,7 @@ @Override public Instruction execute (ThreadInfo ti) { - // Constructors don't return anything so this is the only instruction that can be used to return from a constructor. - - //MethodInfo mi = ti.getMethod(); // Get the current method being executed (e.g. returned from). - if (mi.isInit()) { // Check to see if this method is a constructor. - int objref = ti.getThis(); ElementInfo ei = ti.getElementInfo(objref); // Get the object. @@ -52,6 +47,9 @@ ei.setConstructed(); } } + + } else if (mi.isClinit()) { + mi.getClassInfo().setInitialized(); } return super.execute(ti);