Linux: thread attach/external suspend/vm_suspend race

###@###.### 2002-11-14

Linux stress testing on 2002.11.11 shows a 5% MultiThreadedBreakpoints
test failure with the following assertion:

# assert(jt->has_last_Java_frame(), "must have last Java frame")
#
# Error happened during: thread suspend
#
# Error ID: src/os/linux/vm/os_linux.cpp, 1885

PepTest shows a 2-3% failure rate with the same assertion. There were
also some hangs.

I did some poking around and found a race that explains the assertion
failures. I have not gotten a stack track from the hang yet, but its
frequency is very low.

jni_DestroyJavaVM() creates a new JavaThread via
JavaVM::AttachCurrentThread(). This is implemented by
jni_AttachCurrentThread() which adds the thread to the threads list
before creating the Java level thread object via a call to
JavaThread::allocate_threadObj().

Once the thread is on the threads list, it is possible for the JVM/DI
GetAllThreads() API to see it. Thus, the JPDA backend will have a
chance to externally suspend the thread before its THREAD_START event.

In the case of this thread, it self-suspends in
JavaCallWrapper::JavaCallWrapper() just after changing its
ThreadState from _thread_in_vm to _thread_in_Java. It has not yet
run any Java code so _last_Java_sp is NULL.

The VMThread starts a safepoint at the same time that the thread is
trying to self-suspend. If the VMThread makes it past its
is_any_suspended_with_lock() check, it will:

- vm_suspend() the thread
- find the thread in state _thread_in_Java
- call get_top_frame()
- call fetch_top_frame()
- fail an assertion because _last_Java_sp is NULL

The assertion is new with the Itanium merge which is why I didn't see
this problem with my early Mantis stress runs.