view src/main/gov/nasa/jpf/jvm/bytecode/INVOKECLINIT.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
line wrap: on
line source

/*
 * Copyright (C) 2014, United States Government, as represented by the
 * Administrator of the National Aeronautics and Space Administration.
 * All rights reserved.
 *
 * The Java Pathfinder core (jpf-core) platform is licensed under the
 * Apache License, Version 2.0 (the "License"); you may not use this file except
 * in compliance with the License. You may obtain a copy of the License at
 * 
 *        http://www.apache.org/licenses/LICENSE-2.0. 
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and 
 * limitations under the License.
 */
package gov.nasa.jpf.jvm.bytecode;

import gov.nasa.jpf.vm.ClassInfo;
import gov.nasa.jpf.vm.ElementInfo;
import gov.nasa.jpf.vm.Instruction;
import gov.nasa.jpf.vm.MethodInfo;
import gov.nasa.jpf.vm.ThreadInfo;

/**
 * this is an artificial bytecode that we use to deal with the particularities of 
 * <clinit> calls, which are never in the loaded bytecode but always directly called by
 * the VM. The most obvious difference is that <clinit> execution does not trigger
 * class initialization.
 * A more subtle difference is that we save a wait() - if a class
 * is concurrently initialized, both enter INVOKECLINIT (i.e. compete and sync for/on
 * the class object lock), but once the second thread gets resumed and detects that the
 * class is now initialized (by the first thread), it skips the method execution and
 * returns right away (after deregistering as a lock contender). That's kind of hackish,
 * but we have no method to do the wait in, unless we significantly complicate the
 * direct call stubs, which would obfuscate observability (debugging dynamically
 * generated code isn't very appealing). 
 */
public class INVOKECLINIT extends INVOKESTATIC {

  public INVOKECLINIT (ClassInfo ci){
    super(ci.getSignature(), "<clinit>", "()V");
  }

  @Override
  public Instruction execute (ThreadInfo ti) {    
    MethodInfo callee = getInvokedMethod(ti);
    ClassInfo ciClsObj = callee.getClassInfo();
    ElementInfo ei = ciClsObj.getClassObject();

    if (ciClsObj.isInitialized()) { // somebody might have already done it if this is re-executed
      if (ei.isRegisteredLockContender(ti)){
        ei = ei.getModifiableInstance();
        ei.unregisterLockContender(ti);
      }
      return getNext();
    }

    // not much use to update sharedness, clinits are automatically synchronized
    if (reschedulesLockAcquisition(ti, ei)){     // this blocks or registers as lock contender
      return this;
    }
    
    // if we get here we still have to execute the clinit method
    setupCallee( ti, callee); // this creates, initializes & pushes the callee StackFrame, then acquires the lock
    ciClsObj.setInitializing(ti);

    return ti.getPC(); // we can't just return the first callee insn if a listener throws an exception
  }

  @Override
  public boolean isExtendedInstruction() {
    return true;
  }

  public static final int OPCODE = 256;

  @Override
  public int getByteCode () {
    return OPCODE;
  }
  
  @Override
  public void accept(JVMInstructionVisitor insVisitor) {
	  insVisitor.visit(this);
  }
}