Troubleshooting Guide for Java SE 6 with HotSpot VM

Chapter 5

Troubleshooting Hanging or Looping Processes

This chapter provides information and guidance on some specific procedures for troubleshooting hanging or looping processes.

Problems can occur that involve hanging or looping processes. A hang can occur for many reasons but often stems from a deadlock in application code, API code, or library code. A hang can even be due to a bug in the HotSpot virtual machine.

Sometimes an apparent hang turns out to be, in fact, a loop. For example, a bug in a VM process that causes one or more threads to go into an infinite loop can consume all available CPU cycles.

An initial step when diagnosing a hang is to find out if the VM process is idle or consuming all available CPU cycles. To do this requires using an operating system utility. If the process appears to be busy and is consuming all available CPU cycles then it is likely that the issue is a looping thread rather than a deadlock. On Solaris OS, for example, the command prstat -L -p <pid> can be used to report the statistics for all LWPs in the target process and thus will identify the threads that are consuming a lot of CPU cycles.

5.1 Diagnosing a Looping Process

If a VM process appears to be looping, the first step is to try to get a thread dump. If a thread dump can be obtained, it will often be clear which thread is looping. If the looping thread can be identified, then the trace stack in the thread dump can provide direction on where (and maybe why) the thread is looping.

If the application console (standard input/output) is available, then press the Ctrl-\ key combination (on Solaris OS or Linux) or the Ctrl-Break key combination (on Windows) to cause the HotSpot VM to print a thread dump, including thread state. On Solaris OS and Linux the thread dump can also be obtained by sending a SIGQUIT to the process (command kill -QUIT <pid>). In this case the thread dump is printed to the standard output of the target process. The output might be directed to a file, depending on how the process was started.

If the Java process is started with the -XX:+PrintClassHistogram command-line option, then the Ctrl-Break handler will produce a heap histogram.

If a thread dump can be obtained, then a good place to start is the thread stacks of the threads that are in the runnable state. See 2.15.1 Thread Dump for information on the format of the thread dump, as well as a table of the possible thread states in the thread dump. In some cases it might be necessary to get a sequence of thread dumps in order to determine which threads appear to be continuously busy.

If the application console is not available (process is running as a background process, or the VM output is directed to an unknown location), then the jstack utility can be used to obtain the stack thread. Use the jstack -F pid option to force a stack dump of the looping process. See 2.11 jstack Utility for information on the output of this utility. The jstack utility should also be used if the thread dump does not provide any evidence that a Java thread is looping.

When reviewing the output of the jstack utility, focus initially on the threads that are in the RUNNABLE state. This is the most likely state for threads that are busy and possibly looping. It might be necessary to execute jstack a number of times to get a more complete picture of which threads are looping. If a thread appears to be always in the RUNNABLE state, then the -m option can be used to print the native frames and can provide a further hint on what the thread is doing. If a thread appears to be looping continuously while in the RUNNABLE state, this situation can indicate a potential HotSpot VM bug that needs further investigation.

If the VM does not respond to a Ctrl-\ this could indicate a VM bug rather than an issue with application or library code. In this case use jstack with the -m option (in addition to the -F option) to get a thread stack for all threads. The output will include the thread stacks for VM internal threads. In this stack trace, identify threads that do not appear to be waiting. For example, on Solaris OS you identify the threads that are not in functions such as __lwp_cond_wait, __lwp_park, ___pollsys, or other blocking functions. If it appears that the looping is caused by a VM bug, then collect as much data as possible and submit a bug report. See Chapter 7, Submitting Bug Reports for more details on data collection.

5.2 Diagnosing a Hung Process

If the application appears to be hung and the process appears to be idle, then the first step is to try to obtain a thread dump. If the application console is available, then press the Ctrl-\ keys (on Solaris OS or Linux) or the Ctrl-Break keys (on Windows) to cause the HotSpot VM to print a thread dump. On Solaris OS and Linux the thread dump can also be obtained by sending a SIGQUIT to the process (command kill -QUIT <pid>).

5.2.1 Deadlock Detected

If the hung process is capable of generating a thread dump, then the output is printed to the standard output of the target process. After printing the thread dump, the HotSpot VM executes a deadlock detection algorithm. If a deadlock is detected it will be printed along with the stack trace of the threads involved in the deadlock. Below is an example of this output.

Found one Java-level deadlock:
  waiting to lock monitor 0x000ffbf8 (object 0xf0c30560, a java.awt.Component$AWTTreeLock),
  which is held by "main"
  waiting to lock monitor 0x000ffe38 (object 0xf0c41ec8, a java.util.Vector),
  which is held by "AWT-EventQueue-0"

Java stack information for the threads listed above:
        at java.awt.Container.removeNotify(
        - waiting to lock <0xf0c30560> (a java.awt.Component$AWTTreeLock)
        at java.awt.Window$
        at java.awt.Window.doDispose(
        at java.awt.Dialog.doDispose(
        at java.awt.Window.dispose(
        at java.awt.Window.disposeImpl(
        at java.awt.Window$
        - locked <0xf0c41ec8> (a java.util.Vector)
        at java.awt.Window.doDispose(
        at java.awt.Window.dispose(
        at javax.swing.SwingUtilities$SharedOwnerFrame.dispose(
        at javax.swing.SwingUtilities$SharedOwnerFrame.windowClosed(
        at java.awt.Window.processWindowEvent(
        at javax.swing.JDialog.processWindowEvent(
        at java.awt.Window.processEvent(
        at java.awt.Component.dispatchEventImpl(
        at java.awt.Container.dispatchEventImpl(
        at java.awt.Window.dispatchEventImpl(
        at java.awt.Component.dispatchEvent(
        at java.awt.EventQueue.dispatchEvent(
        at java.awt.EventDispatchThread.pumpOneEventForHierarchy(
        at java.awt.EventDispatchThread.pumpEventsForHierarchy(
        at java.awt.EventDispatchThread.pumpEvents(
        at java.awt.EventDispatchThread.pumpEvents(
        at java.awt.Window.getOwnedWindows(
        - waiting to lock <0xf0c41ec8> (a java.util.Vector)
        at javax.swing.SwingUtilities$SharedOwnerFrame.installListeners(
        at javax.swing.SwingUtilities$SharedOwnerFrame.addNotify(
        at java.awt.Dialog.addNotify(
        - locked <0xf0c30560> (a java.awt.Component$AWTTreeLock)
        at java.awt.Dialog.conditionalShow(
        - locked <0xf0c30560> (a java.awt.Component$AWTTreeLock)
        at java.awt.Component.setVisible(
        at test01.main(

Found 1 deadlock.

The default deadlock detection works with locks that are obtained using the synchronized keyword, as well as with locks that are obtained using the java.util.concurrent package. If the Java VM flag -XX:+PrintConcurrentLocks is set, then the stack trace also shows a list of lock owners.

If deadlock is detected, then you must examine the output in more detail in order to understand the deadlock. In the above example the thread main is locking object <0xf0c30560> and is waiting to enter 0xf0c41ec8, which is locked by thread AWT-EventQueue-0. However, thread AWT-EventQueue-0 is waiting to enter 0xf0c30560, which is locked by main.

The detail in the stack traces provides information to help find the deadlock.

5.2.2 Deadlock Not Detected

If the thread dump is printed and no deadlocks are found, then the issue might be a bug in which a thread waiting on a monitor that is never notified. This could be a timing issue or a general logic bug.

To find out more about the issue, examine each of the threads in the thread dump and each thread that is blocked in Object.wait(). The caller frame in the stack trace indicates the class and method that is invoking the wait() method. If the code was compiled with line number information (the default), then this provides direction as to the code to examine. In most cases you must have some knowledge of the application logic or library in order to diagnose this issue further. In general you must understand how the synchronization works in the application and in particular the details and conditions for when and where monitors are notified.

5.2.3 No Thread Dump

If the VM does not respond to a Ctrl-\ or Ctrl-Break, then it is possible that the VM is deadlocked or hung for some other reason. In that case use the jstack utility (see 2.11 jstack Utility ) to obtain a thread dump. Use the jstack -F pid option to force a stack dump of the hung process. This also applies in the case where the application is not accessible or the output is directed to an unknown location.

In the jstack output, examine each of the threads in the BLOCKED state. The top frame can sometimes indicate why the thread is blocked, for example, Object.wait or Thread.sleep. The rest of the stack will give an indication of what the thread is doing. This is particularly true when the source has been compiled with line number information (the default) and you can cross reference the source code.

If a thread is in the BLOCKED state and the reason is not clear, then use the -m option to get a mixed stack. With the mixed stack output, it should be possible to identify why the thread is blocked. If a thread is blocked trying to enter a synchronized method or block, then you will see frames such as ObjectMonitor::enter near the top of the stack. Below is an example.

----------------- t@13 -----------------
0xff31e8b8      ___lwp_cond_wait + 0x4
0xfea8c810      void ObjectMonitor::EnterI(Thread*) + 0x2b8
0xfeac86b8      void ObjectMonitor::enter2(Thread*) + 0x250

Threads in the RUNNABLE state might also be blocked. The top frames in the mixed stack should indicate what the thread is doing.

One specific thread to check is VMThread. This is the special thread used to execute operations like garbage collection. It can be identified as the thread that is executing VMThread::run() in its initial frames. On Solaris OS it is typically t@4. On Linux it should be identifiable using the C++ mangled name _ZN8VMThread4loopEv.

In general the VM thread is in one of three states: waiting to execute a VM operation, synchronizing all threads in preparation for a VM operation, or executing a VM operation. If you suspect that a hang is a HotSpot VM bug rather than an application or class library deadlock, then pay special attention to the VM thread.

If the VM thread appears to be stuck in SafepointSynchronize::begin, then this could indicate an issue bringing the VM to a safepoint. A safepoint indicates that all threads executing in the VM are blocked and waiting for a special operation, such as garbage collection, to complete.

If the VM thread appears to be stuck in function, where function ends in doit, then this could also indicate a VM problem.

In general, if you can execute the application from the command line, and you get to a state where the VM does not respond to a Ctrl-\ or Ctrl-Break, it is more likely that you have uncovered a VM bug, a thread library issue, or a bug in another library. If this occurs, obtain a crash dump (see 7.4 Collecting Core Dumps for instructions on how to do this), gather as much information as possible, and submit a bug report or support call.

One other tool to mention in the context of hung processes is the pstack utility on Solaris OS. On Solaris 8 and 9 OS, this utility prints the thread stacks for LWPs in the target process. On Solaris 10 OS and starting with the JDK 5.0 release, the output of pstack is similar, though not identical, to the output from jstack -m. As with jstack, the Solaris 10 OS implementation of pstack prints the fully qualified class name, method name, and bci. It will also print line numbers for the cases where the source was compiled with line number information (the default). This is useful for developers and administrators who are familiar with the other utilities on Solaris OS that exercise features of the /proc file system.

The equivalent tool of pstack on Linux is lsstack. This utility is included in some distributions and otherwise obtained from the web site. At the time of this writing, lsstack reported native frames only.

5.3 Solaris 8 OS Thread Library

The default thread library on Solaris 8 OS is often referred to as the T1 library. This thread library implemented the m: n threading model, where m user threads are mapped to n kernel-level threads (LWPs). Solaris 8 OS also shipped with an alternative and newer thread library in /usr/lib/lwp. The alternative thread library is often referred to as the T2 library, and it became the default thread library in Solaris 9 and 10 OS. In older releases of J2SE (pre-1.4.0 in particular) there were a number of issues with the default thread library, for example, bugs in the thread library, LWP synchronization problems, or LWP starvation. LWP starvation is a scenario in which there are user threads in the runnable state but there are no kernel level threads available.

Although the issues cited are historical, it should be noted that when the JDK software is deployed on Solaris 8 OS, it still uses the T1 library by default. LWP starvation type issues do not arise because the JDK release uses “bound threads” so that each user thread is bound to a kernel thread. However in the event that you encounter an issue, such as a hang, which you believe is a thread library issue, then you can instruct the HotSpot VM to use the T2 library by adding /usr/lib/lwp to the LD_LIBRARY_PATH. To check if the T2 library is in use, issue the command pldd <pid> to list the libraries loaded by the specified process.