Name: dd4877 Date: 10/03/2001


daniel.daugherty@sun 2001-10-03

One of my profiling stress runs crashed on the following assertion
failure:

This run crashed on the following assertion failure:

# assert(state == _thread_in_Java || state == _thread_new_trans
# || state == _thread_in_native || state == _thread_in_native_trans,
# "a suspended thread should have thread state in_native, in_Java, or just start
ed")
#
# Error ID: baseline-stress/src/share/vm/runtime/safepoint.cpp, 438 [ Patched ]

While the VMThread was calling SafepointSynchronize::begin(),
examine_state_of_thread() found a suspended thread with an unexpected
thread state value. This was caused by a race between JNI
DestroyJavaVM() and safepointing.

Here is the scenario:

- JNI DestroyJavaVM() has been called
  - it queues up a VM_SuspendDaemons operation
  - it is waiting in VMThread::wait_for_vm_thread_exit()
- VMThread
  - processes the VM_SuspendDaemons operation which is defined as an
    "evaluate at safepoint" operation
- JavaThread (daemon)
  - coming out of an Object.wait()
  - sees that there is a pending safepoint
  - blocks in the thread state transition from _thread_blocked to
    _thread_in_vm (state now _thread_blocked_trans)
  - gets suspended by the VMThread
- VMThread
  - does another safepoint (we have SafepointALot set)
  - fails the assertion

======================================================================