Don't Hesitate to Terminate

From EDM2
Jump to: navigation, search

By Monte Copeland

Applications, unlike operating systems, have to terminate. As programmers, we concentrate on the loading and running of our applications, but we often skimp on termination. As a program grows in complexity, with multiple processes, threads, DLLs, semaphores, and shared memory, termination becomes a complex issue. This article discusses several aspects of termination for programs running under OS/2.

Thread 1 Is Special

Under OS/2, thread 1 has special character. When an EXE loads, OS/2 creates thread 1 in the new process and gives it control. Thread 1 may create other threads, starting with thread 2. There is no thread 0 in OS/2.

No matter how many threads are in a process, the process will terminate when thread 1 terminates. In other words, if thread 1 calls

rc = DosExit( EXIT_THREAD, 0 );

it is practically the same as

rc = DosExit( EXIT_PROCESS, 0 );

In either case, thread 1 receives the exception XCPT_PROCESS_TERMINATE. Any other threads receive the exception XCPT_ASYNC_PROCESS_TERMINATE.

When a thread receives either of these exceptions, it should remove its exception handler, clean up, and exit. It can exit by calling DosExit( <thread>, 0 ) or by returning from its thread function. The latter is better because it gives the language runtime a chance to clean up, too.

The sample code T1TERM.C on the DevCon OS/2 disc 1 in \SOURCE\DEVNEWS\VOL12\TERM demonstrates this behavior and the associated exceptions.

DosKillThread() API

A hot topic on OS/2 programming forums is the use of DosKillThread(). OS/2 programmers have tried to use this API and sometimes found it troublesome. DosKillThread() can work correctly in some cases; in other cases it hangs the system. Instead of using DosKillThread(), programmers usually set a flag so the thread exits itself.

DosKillThread() is troublesome because most code is not robust with regard to killed threads. Killing a thread is a random act. Robust code is always prepared for this asynchronous exception, and robustness requires tireless use of exception handlers.

DosKillThread() causes the most trouble in Presentation Manager applications or applications using 16-bit DLLs. For performance reasons, PM does not use exception handlers to guard against the XCPT_ASYNC_PROCESS_TERMINATE exception. So if a thread is killed while it owns the main semaphore of PM, PM can hang solidly. A similar problem exists for threads blocked in 16-bit DLLs. They don't have exception handlers because DosSetExceptionHandler() is a 32-bit API.

However, DosKillThread() works well on threads blocked in OS/2 kernel APIs such as DosWaitEventSem(), DosRequestMutexSem(), DosRead(), or DosSleep(), especially if these APIs were called directly from application code, not indirectly from other DLLs.

See the sample code KILL2.C on the DevCon OS/2 disc 1 in \SOURCE\DEVNEWS\VOL12\TERM. Every subroutine called on thread 2 uses an exception handler to detect the killed-thread case. Once handled, each subroutine raises the same exception again for callers' exception handlers.

Signals Are Exceptions, Too

OS/2 delivers signals as exceptions. Signals result from the user pressing Ctrl-C or Ctrl-Break or because a program called DosKillProcess(). In all cases, OS/2 dispatches signal exceptions on thread 1 of the process.

If a program handles the thread-granular termination exceptions XCPT_PROCESS_TERMINATE and XCPT_ASYNC_PROCESS_TERMINATE, then it is not necessary to handle the signal exception at all. The default action for a signal exception is to terminate all the threads in the process. Thus, every thread in the process will soon receive an XCPT_ASYNC_PROCESS_TERMINATE exception.

Exit Lists Are History

OS/2 maintains a per-process list of functions to be called at process termination to do clean up. Programs add functions to this list using the DosExitList() API. At process termination, OS/2 kills all the process' threads except thread 1. The system uses the remaining thread to call each function in the exit list. Then the process dies.

There are restrictions on exit list functions. Some APIs can't be called from an exit list function, (for example SQL). And don't risk blocking on semaphores while in exit list, or the process will stay blocked until reboot.

Exit lists are rooted in 16-bit OS/2 programming. In 32-bit OS/2, new APIs provide a superset of function which can eliminate exit lists altogether.

First of all, there is per-thread exception handling to detect when individual threads terminate. (See the sample code KILL2.C or T1TERM.C mentioned above. Refer to Volume 2 of The Developer Connection News online (\DOCS\DEVCON2.INF) for an article on exception management.) This mechanism provides thread-granular clean up instead of process-granular clean up.

Second, the DosWaitThread() API lets the programmer ensure that thread 1 is the last thread to exit. This is important for the reason stated earlier: when thread 1 exits, all threads exit.

For DLLs, the IBM VisualAge C++ compiler documents the special function:

ULONG APIENTRY _DLL_InitTerm( ULONG hmod, ULONG flag );

which is called during process initialization and termination, assuming the DLL is linked with INITINSTANCE and TERMINSTANCE attributes on the LIBRARY statement in the .DEF file. When called at termination time, it is not so different from an exit list function call: there is only thread 1 left in the process, and the process is about to die.

PM Application Termination

Presentation Manager applications begin to terminate upon receipt of the WM_CLOSE message. Default WM_CLOSE processing posts a WM_QUIT message to the same window, and this causes the get/dispatch message loop to exit. After that, the application destroys its windows and cleans up.

WM_CLOSE message processing IS NOT a good place to perform application shutdown because PM applications are not guaranteed to receive a WM_CLOSE. For example, a user can terminate a PM application by pressing the Delete key at the Window List (or mouse button 2 then the close option). The shell does not send a WM_CLOSE message to this application. Instead it posts a WM_QUIT message to the message queue of the application.

Note: For best results, put application termination and clean up code after the get/dispatch message loop, not in the WM_CLOSE case. PM applications can handle the XCPT_SIGNAL exception and still terminate in their usual way. See sample code PMAPP.C on disc 1 of your Developer Connection for OS/2 CD-ROM. Start PMAPP.EXE then kill it with KILL.EXE. Upon receipt of the XCPT_SIGNAL exception, PMAPP.EXE posts itself a WM_QUIT message and terminates in its normal way.

OS/2 Cleans Up At Termination

Neatness counts at program termination. Of course it is best for an application to clean up its files, memory, semaphores, and windows before termination. However, OS/2 will clean up at process termination in case of abnormal end or laziness.

At process termination, OS/2 closes open files, closes private semaphores, and frees private memory. OS/2 decrements the use count for shared semaphores and closes them if the use count goes to zero. OS/2 decrements the use count for shared memory and frees the memory if the use count goes to zero. At thread termination, OS/2 frees the stack of the thread, provided it was created with 32-bit _beginthread() or 32-bit DosCreateThread().

For untidy PM applications, the OS/2 window manager will clean up windows and messages queues, and the graphics engine will close device contexts and presentation spaces.

REXX Terminates

Any application can run REXX programs using the RexxStart() API to give a compiled application the flavor of an interpreter. A thread in a process can halt another of its threads executing REXX code using the RexxSetHalt() API. However, if the REXX thread is blocked in OS/2 (waiting on a semaphore or I/O, for example), RexxSetHalt() will not interrupt it. You must combine RexxSetHalt() with some form of wake up. In your REXX command handlers and registered functions, block REXX threads in muxwait event semaphores then post one after the call to RexxSetHalt().

Session Manager Reports Terminations

The OS/2 session manager knows when a related child session terminates and can notify the parent session via a queue. This is a nice feature for a writing specialized shells. One shell process can start a number of child sessions all with the same termination queue and devote one thread to read it. When any child session terminates, the thread blocked in DosReadQueue() will wake up. See \SOURCE\DEVNEWS\VOL12\TERM\TERMQ.C on disc 1 for some sample code.

Termination and Shutdown

System administrators always need to shut down OS/2 computers for maintenance. They need remote shutdown that is safe and automatic. It sounds simple, but it's BIG trouble when applications refuse to terminate.

SETBOOT.EXE of OS/2 is a limited-function program that simulates Ctrl-Alt-Del for reboot. Before reboot, you must always close open files and terminate important programs.

For example, termination of an OS/2 server is the reverse of startup:

  • stop clients using server programs and resources
  • stop server applications and daemons
  • stop DB2
  • stop OS/2 Warp Server
  • stop MQSeries
  • stop Communications Manager
  • stop remaining programs that have open files
  • setboot.exe

Besides passing parameters to Boot Manager, SETBOOT.EXE does little more than emulate Ctrl-Alt-Del.

WARNING
Setboot alone is not enough to properly shutdown a machine. True, the file system is shut down, and there's no CHKDSK at next IPL, but any read/write files left open during setboot are corrupted.

Some software products provide a program to stop a program, a utility EXE that signals the main EXE(s) to terminate. This is convenient and helps greatly to automate unattended shutdowns. Ideally, such a utility would terminate after the main program terminates thus eliminating guesswork about the timing of termination.

So Don't Hesitate to Terminate

We work hard to get our programs up and running, but don't forget about termination. Write programs so they terminate quickly and cleanly at the right time. Your users and system's manager will love you for it.

Reprint Courtesy of International Business Machines Corporation, © International Business Machines Corporation