Uploaded image for project: 'JDK'
  1. JDK
  2. JDK-5063437

JVMPI obj_frees come before method_exits

    XMLWordPrintable

    Details

    • Subcomponent:
    • CPU:
      sparc
    • OS:
      solaris_8

      Backports

        Description

        Name: dk106046 Date: 06/15/2004

        OPERATING SYSTEM(S): Solaris 5.8

        FULL JDK VERSION(S):
        1.3.1_12-b03


        1) Set the path to the java.
        2) Unzip the jdk13bug.zip onto any directory (say /usr/tmp)
        3) Ensure tha the directory containing the libmyprofiler.so is on the LD_LIBRARY_PATH
        ie
        LD_LIBRARY_PATH=/usr/tmp/jdk13bug/solaris:$LD_LIBRARY_PATH
        export LD_LIBRARY_PATH
        4) Run the testcase
        java -Xrunmyprofiler PerformanceExample

        In our profiler we enable JVMPI_EVENT_METHOD_ENTRY2 and JVMPI_EVENT_METHOD_EXIT, and keep
        track of the method stack that we have seen so far. That is, every time we see a
        JVMPI_EVENT_METHOD_ENTRY2 we push an item onto the method stack, and every time we see
        JVMPI_EVENT_METHOD_EXIT, we pop an item from the method stack. This stack is kept track
        of on a per-thread basis.
        The problem we have been seeing is that after we receive a JVMPI_EVENT_METHOD_ENTRY2
        (referencing an object in event->u.method_entry2.obj_id) we will subsequently see a
        JVMPI_EVENT_OBJECT_FREE event on that obj_id BEFORE the corresponding
        JVMPI_EVENT_METHOD_EXIT has occurred. That is, references to obj_id are still on the
        method stack when the OBJECT_FREE event occurs. This doesn't make sense. If the JVM is
        still executing some method on object obj_id, then it cannot free that object. Hence, either
        the OBJECT_FREE event is erroneous, or METHOD_EXIT events are being lost somewhere.
        The 'test' profiler (libmyprofiler.so) to replicate this problem is based upon
        the jvmpi_example profiler available from the JVMPI spec website.
        It's behaviour is as follows: It keeps track of the method stack on a per-thread basis.
        Whenever a JVMPI_EVENT_METHOD_ENTRY2 occurs on some thread x, it pushes
        event->u.method_entry2.obj_id onto the stack associated with thread x; and whenever a
        JVMPI_EVENT_METHOD_EXIT occurs on thread x, it pops the stack associated with thread x.
        Whenever a JVMPI_EVENT_OBJ_FREE event occurs, we check to make sure that the object being
        freed is not referenced on any of the method stacks. If it is, an error message is printed:
        ERROR> Freeing Object currently on a method stack!!!!
        (Note, in this test profiler, we also handle JVMPI_EVENT_OBJECT_MOVE's if they occur, so
        that all obj_id references in all the method stacks are updated as necessary).

        This problem does not occur with the latest Windows JVMs on 1.3.1. A testcase on Windows can be
        provided. This does not occur on 1.4.x and above, even on Solaris.

        ======================================================================
        ###@###.### 10/14/04 12:57 GMT

          Attachments

            Issue Links

              Activity

                People

                Assignee:
                mbankal Mala Bankal (Inactive)
                Reporter:
                dkorbel David Korbel (Inactive)
                Votes:
                0 Vote for this issue
                Watchers:
                0 Start watching this issue

                  Dates

                  Created:
                  Updated:
                  Resolved:
                  Imported:
                  Indexed: