Advanced Guide to Hang Analysis

What is a hang?

It's an external symptom or a user perception that little or no work is being done. It could be a case of extremely poor performance. The hang symptom categorises itself into three distinct cases:

Wait

Threads and processes are not being dispatched by the Operatin System. Thread status gives the initial clue. Use of the .P command determines status.

Blocking

Threads may wait voluntarily for a resource or an event, in which case they will probably be Blocked. They might poll for the resource, in which cases they will cycle through blocking, being ready and running until the resource becomes available.

A notoriously problematic case of Blocking is the deadlock. This is where two threads are each own exclusively a resource and block waiting for ownership of the other's resource.

Suspension
Another thread may have deliberately debarred a thread from being scheduled, in which case we will see the waiting thread in a crt or frz state.
Pre-emption
Another thread monopolises the system. Typically the hanging threads will be ready for dispatch (rdy), but will never or rarely receive a minimum time-slice. We look for running and ready threads with an excessively high priority. The .P will give the calculated priority of each thread.
Disabled wait
Looks rather like a H/W Freeze. Last instruction executed was a HLT having sometime previously disabled interrupts using CLI. This usually happens only when ring 0 code detects an terminal condition. One would hope that some form of diagnostic information had been displayed prior to this particularly if NMIs have been disabled also! If NMIs have not been disabled then an artificially generated channel check may be used to cause an NMI, which would allow one to break into the kernel debugger. However the KDB to allows only one NMI channel check per boot. If NMIs are disabled then H/W analysis techniques may be the only recourse.
Loop
A thread is running more or less permanently. It's state will mostly be rdy or run. Similar analysis techniques to trap apply here. We examine the registers of the running thread by using .R. From the we can determine in which module the thread is running by looking at the owner of the executing code segment. If necessary we unwind the stack and determine the caller etc.. Analysis is very similar to trap analysis.
H/W Freeze.
The processor fetch-execute cycle has been suspended. Not even an NMI interrupt will resume instruction fetch. This is almost certainly a hardware problem. Timing/clocking problems caused by incompatible cards, overloaded busses, incorrect bus terminations, faulty processor or controller chips. Use H/W techniques such an ICE machine or Logic Analyser.

The following Theory Topics are now covered in detail:

• The Wait Condition - Diagnostic Techniques

  This is further subdivided into two topics of discussion:

  Memory Management and Ownership Topics

  Thread Scheduling and Dispatching

  
  Program Design Issues and Weaknesses

  The final section of this Guide is a collection of Worked Examples that explore memory management, the File System, Presentation Manager and Ring 0 Loops From a Dump.

  ---

  [Back: A Form to Use For Unwinding Stacks]
  [Next: The Wait Condition - Diagnostic Techniques]