diff -r 8ebcedb7604d -r 716c64bda5ba src/share/vm/runtime/thread.cpp
--- a/src/share/vm/runtime/thread.cpp	Thu Oct 18 13:09:47 2012 -0400
+++ b/src/share/vm/runtime/thread.cpp	Fri Oct 19 21:40:07 2012 -0400
@@ -323,12 +323,10 @@
   os::initialize_thread(this);
 
 #if INCLUDE_NMT
-   // record thread's native stack, stack grows downward
-  if (MemTracker::is_on()) {
-    address stack_low_addr = stack_base() - stack_size();
-    MemTracker::record_thread_stack(stack_low_addr, stack_size(), this,
+  // record thread's native stack, stack grows downward
+  address stack_low_addr = stack_base() - stack_size();
+  MemTracker::record_thread_stack(stack_low_addr, stack_size(), this,
       CURRENT_PC);
-  }
 #endif // INCLUDE_NMT
 }
 
@@ -345,6 +343,9 @@
   if (_stack_base != NULL) {
     address low_stack_addr = stack_base() - stack_size();
     MemTracker::release_thread_stack(low_stack_addr, stack_size(), this);
+#ifdef ASSERT
+    set_stack_base(NULL);
+#endif
   }
 #endif // INCLUDE_NMT
 
@@ -1521,10 +1522,12 @@
       tty->print_cr("terminate thread %p", this);
   }
 
-  // Info NMT that this JavaThread is exiting, its memory
-  // recorder should be collected
+  // By now, this thread should already be invisible to safepoint,
+  // and its per-thread recorder also collected.
   assert(!is_safepoint_visible(), "wrong state");
-  MemTracker::thread_exiting(this);
+#if INCLUDE_NMT
+  assert(get_recorder() == NULL, "Already collected");
+#endif // INCLUDE_NMT
 
   // JSR166 -- return the parker to the free list
   Parker::Release(_parker);
@@ -2425,6 +2428,7 @@
 }
 
 void JavaThread::remove_stack_guard_pages() {
+  assert(Thread::current() == this, "from different thread");
   if (_stack_guard_state == stack_guard_unused) return;
   address low_addr = stack_base() - stack_size();
   size_t len = (StackYellowPages + StackRedPages) * os::vm_page_size();
@@ -4093,7 +4097,10 @@
 
     // Now, this thread is not visible to safepoint
     p->set_safepoint_visible(false);
-
+    // once the thread becomes safepoint invisible, we can not use its per-thread
+    // recorder. And Threads::do_threads() no longer walks this thread, so we have
+    // to release its per-thread recorder here.
+    MemTracker::thread_exiting(p);
   } // unlock Threads_lock
 
   // Since Events::log uses a lock, we grab it outside the Threads_lock