###@###.### 2003-06-19

Here is the first variation of this deadlock:

The Java2Demo program is done and has called System.exit(0).
The JVM_Halt thread calls before_exit() and then requests a
VM_Exit operation. The VMThread is trying to execute a
safepoint. A JavaThread is executing a monitorexit bytecode and
receives a SIGPROF. This JavaThread is still "running".
The VMThread is in SafepointSynchronize::begin() in the
'while(still_running > 0)' loop waiting for the JavaThread to
stop running.

The SIGPROF handler calls JVM/PI GetThreadLocalStorage() and
because of the before_exit() call, the real handler for the API
has been replaced with a call to _block().

Here is a stack trace snippet from the JavaThread that called
_block():

THREAD t@41

  [6] os::yield_all(attempts = 100), line 2046 in "os_solaris.cpp"
  [7] _block(), line 176 in "jvmpi.cpp"
  [8] profhandler(sig = 29, siginfo = 0xd1d3fa38, ucontext = 0xd1d3f838), line 1
31 in "b4757672.c"
  [9] __sighndlr(0x1d, 0xd1d3fa38, 0xd1d3f838, 0xdcbe1008), at 0xdfb92963
  [10] call_user_handler(0x1d, 0xd1d3fa38, 0xd1d3f838), at 0xdfb8e0b6
  [11] sigacthandler(), at 0xdfb8e1e0
  ---- called from signal handler with signal 29 (SIGPROF) ------
  [12] __lwp_mutex_unlock(0x8127f54), at 0xdfaeb80c
  [13] os::Solaris::mutex_unlock(mx = 0x8127f54), line 178 in "os_solaris.hpp"
  [14] os::Solaris::OSMutex::unlock(this = 0x8127f48), line 515 in "os_solaris.h
pp"
  [15] ObjectMonitor::exit(this = 0x8127f2c, __the_thread__ = 0x85c0b58), line 2
80 in "objectMonitor_solaris.cpp"
  [16] ObjectSynchronizer::slow_exit(object = 0xd28ef798, lock = 0xd1d3fbc4, __t
he_thread__ = 0x85c0b58), line 351 in "synchronizer.cpp"
  [17] Runtime1::monitorexit(thread = 0x85c0b58, lock = 0xd1d3fbc4), line 907 in
 "c1_Runtime1.cpp"

The complete stack trace is attached as threads.log.41.


Here is the second variation of this deadlock:

The CPUSampling/PepTest program is done and main() is
returning, i.e., no System.exit() call. The JVM DestroyJavaVM
thread has called before_exit() and is waiting for the VMThread
to exit. The VMThread is trying to execute a safepoint. A
CompilerThread is trying alloc memory and receives SIGPROF.
This CompilerThread has been marked with "callback". VMThread
is in SafepointSynchronize::begin() in the 'while
(_waiting_to_block > 0)' loop waiting for the CompilerThread to
call back.

The SIGPROF handler calls JVM/PI GetThreadLocalStorage() and
because of the before_exit() call, the real handler for the API
has been replaced with a call to _block().

Here is a stack trace snippet from the JavaThread that called
_block():

THREAD t@8

  [6] os::yield_all(attempts = 100), line 2055 in "os_solaris.cpp"
  [7] _block(), line 183 in "jvmpi.cpp"
  [8] profhandler(sig = 29, siginfo = 0xd9feeab0, ucontext = 0xd9fee8b0), line 1
31 in "b4757672.c"
  [9] __sighndlr(0x1d, 0xd9feeab0, 0xd9fee8b0, 0xdc971008), at 0xdfb92963
  [10] call_user_handler(0x1d, 0xd9feeab0, 0xd9fee8b0), at 0xdfb8e0b6
  [11] sigacthandler(), at 0xdfb8e1e0
  ---- called from signal handler with signal 29 (SIGPROF) ------
  [12] CodeHeap::verify(this = 0x8067820), line 421 in "heap.cpp"
  [13] CodeHeap::allocate(this = 0x8067820, size = 1400U), line 133 in "heap.cpp
"
  [14] CodeCache::allocate(size = 1400), line 126 in "codeCache.cpp"
  [15] BufferBlob::operator new(s = 48U, size = 1400U), line 209 in "codeBlob.cp
p"
  [16] BufferBlob::create(buffer_size = 1344, name = 0xdf5d1141 "CodeBuffer cons
tructor"), line 204 in "codeBlob.cpp"

The complete stack trace is attached as threads.log.3978.