SYNOPSIS

#include <ggi/gg.h>
typedef void (ggcleanup_func)(void *);
int ggRegisterCleanup(ggcleanup_func *func, void *arg);
int ggUnregisterCleanup(ggcleanup_func *func, void *arg);
void ggCleanupForceExit(void);

DESCRIPTION

ggRegisterCleanup registers a callback function ("handler") to be executed when any abnormal or unexpected program termination is imminent. The function will be called with its argument.

ggUnregisterCleanup cancels a callback installed with ggRegisterCleanup. If more than one exactly identical callbacks have been installed, the most recently installed one is canceled.

ggCleanupForceExit may only be called from within a LibGG cleanup handler. Once ggCleanupForceExit is called, _exit(2) will be explicitly called after all registered cleanup callbacks have completed by ggExit(3), assuming there is no error that prevents them from completing. It is not possible to cancel such a request once it has been made.

Cleanup functions are executed in LIFO order. They are guaranteed to only be executed once during program termination or during ggExit(3).

These functions are for emergency use only, and should not be used as a substitute to proper memory deallocation routines. They should only be used to restore system state that would otherwise be left corrupted after an abnormal program termination, for example, a video-card timing mode or tty mode. When normal termination occurs, ggUnregisterCleanup should be called to systematically remove the emergency callbacks before ggExit(3) or exit(3) are called.

Callback functions registered with ggRegisterCleanup should not themselves invoke (or invoke any subroutines that may in turn invoke) any of the LibGG locking functions ggLockCreate(3), ggLockDestroy(3), ggLock(3), ggUnlock(3), or ggTryLock(3). However, since callbacks are only invoked during emergencies, they should be ignoring locks in general. If a callback function may be used by anything other than LibGG, it must also be reentrant. Callback functions can come at any time, so write them with this in mind -- make them minimal and tolerant of concurrent access to global read/write data and avoid accessing such data in the first place if it is not absolutely necessary.

The callback functions may be called by a normal call to ggExit(3). As such, it is considered best practice to use ggUnregisterCleanup to remove cleanups when gracefully deinitializing LibGG or a library that uses LibGG, before ggExit(3) is called.

ggRegisterCleanup and ggUnregisterCleanup are threadsafe, however, they are not safe to call from a thread that may be canceled during their execution, and they are not guaranteed to be safe to call from LibGG tasks or any other special context such as a signal handler or a asyncronous procedure call. Above all they should not be called from inside a LibGG cleanup handler.

RETURN VALUE

ggRegisterCleanup returns GGI_OK on success, or one of these error codes:

• GGI_EUNKNOWN;
• GGI_ENOMEM;
• GGI_EBUSY.

ggUnregisterCleanup returns GGI_OK on success, or one of these error codes:

• GGI_EBUSY;
• GGI_ENOTALLOC;
• GGI_ENOMEM.

INTERACTION WITH UNIX SIGNALS

On UNIX systems the LibGG cleanup facilities install signal handlers as per signal(3) or sigaction(2). It is advisible to use LibGG cleanup handlers instead of UNIX signals for the purpose of catching fatal signals because they are implemented portably, however, this is not always an option when mixing LibGG with other libraries. When LibGG must be used with other code that also installs signal handlers, consult the following section.

LibGG installs signal handlers for those signals which are normally fatal to the program. The exact set of functions that is caught depends on the software platform. LibGG installs signal handlers when the first LibGG cleanup handler is installed. These may in fact be installed in ggInit(3) as LibGG may use cleanups internally. The only way to be sure that the LibGG signal handlers are installed is to install a cleanup after ggInit(3).

By setting any signal handler to SIG_IGN before calling ggInit(3), the application can force LibGG to ignore the signal, so that cleanups are not run when that particular signal is received. LibGG will also overload any application signal handlers for fatal signals that are present when it installs a signal handler. Overloaded signal handlers will be run before cleanups are run when the signal occurs. The overloaded signal handler is not guaranteed to be called exactly as it would be by the main application, for example, support for the long form of the signal handler prototype available through sigaction(2) on some systems is not yet implemented, though it may be in the future.

If the system uses the tradition (broken) UNIX signal behavior where signal handlers are set to SIG_DFL to 'block' additional occurances, this may result in rare instances where signal handlers are reentered. Thus the signal handler for a given signal may call the overloaded cleanup multiple times before cleanup functions are called. This will also apply to signal handlers that manipulate the signal mask on more advanced sigaction(2) based systems.

Signals that arrive after that signal which triggers the cleanup callbacks may have their handlers run before or during the execution of the cleanups.

After ggInit(3) and ggRegisterCleanup are called, individual signals may be overridden by application-specific signal handlers or set to SIG_IGN or SIG_DFL. This will prevent LibGG cleanups from being run when the signal occurs. Within limits, it is also acceptable to overload the LibGG signal handler. However it is not acceptable to call the LibGG signal handler with signum equal to any signal on which LibGG did not initially install the handler function, and deinitializing LibGG while overloading its signal handler may cause undefined bahavior.

SHORT TUTORIAL ON WRITING DECENT SIGNAL HANDLERS

LibGG attempts to be tolerant of badly written signal handlers, but consistant, correct behavior can only be guaranteed if signal handlers are written within the following guidelines. First, a signal handler must be written using sigaction if sigaction is available, or the results may not be perfect. If sigaction is not available, the proper code for a simple, overloadable signal handler is such:

lasthandler = signal(signum, SIG_DFL));
/* do stuff */
if (lasthandler == SIG_DFL) signal(signum, current_handler);
else signal(signum, lasthandler);

This code looks circumlocuitous but each part is important for maintaining overloadability in a portable fashion. The signal handler should not reinstall itself unless it detects original UNIX signals are in effect by detecting the automatically installed SIG_DFL, or else it might get called directly, skipping the parent signal handler. Installing SIG_DFL temporarily should be harmless to BSD style signals because the OS is required to block stacked signals through some other mechanism until the signal handler returns.

In order to overload a signal handler, again still dealing with the situation where sigaction is not available, the above code can be modified as such:

lasthandler = signal(signum, SIG_DFL);
/* do stuff */
if (had_oldhandler) {
     signal(signum, lasthandler);
     oldhandler(signum);
}
if (lasthandler == SIG_DFL) signal(signum, current_handler);
else signal(signum, lasthandler);

This is not perfect because it may allow lasthandler to be reentered when used on a system with the original UNIX behavior, in the short period between when lasthandler is reinstalled and the oldhandler installs SIG_DFL. However, if the handlers are all reentrant this should work fine. In the BSD behavior, this again is harmless because other OS mechanisms prevent reentry.

Systems without sigaction are pretty cretinous and rarer these days, however. When sigaction(2) is available we can assume that signal handlers do not need to reinstall themselves as per the original UNIX SIG_DFL behavior. As such no special consideration is needed to write a proper overloading/overloadable handler, however, in order to assure that cleanup functions are only run once even in multithreaded, multiprocessor environments, LibGG may need to temporarily overload a signal handler which has overloaded LibGG's signal handler with a dummy pass-through handler, and as of this writing LibGG's behavior when the signal mask is altered is not yet specified and should be considered undefined.

Measures are taken within LibGG to limit the impact of interaction with badly written signal handlers that reinstall their own handler when it is not needed or desired, however it is recommended that libraries that use such handlers be updated to use better code when compiled on more modern systems.

One last note on stacking signal handlers: When writing for an environment where different libraries may overload signals, all libraries must prevent loops from forming. It is not sufficient that they simply check that they never overload their own signal handler, because another library may have overloaded it already, and thus you may have handler A calling handler B calling handler A which then calls handler B again. Libraries must keep track of whether their signal handlers are installed or not through other means.

FALLBACK MODE

LibGG expects some sort of signal-like system to be present in the environment, otherwise there is no way to implement the behavior described above. When LibGG is compiled on a system that has no such support, a fallback mode is invoked where cleanup handlers are registered with the atexit(3) facility, or anything it may have that is like atexit(3). It may not be possible to unregister cleanups supported in such a way, and they will always run at normal program exit, even after LibGG is exited. There is no way for them to run during abnormal termination.