setjmp.h

The example below shows the basic idea of setjmp. There, main() calls first(), which in turn calls second(). Then, second() jumps back into main(), skipping first()'s call of printf().

#include <stdio.h>
#include <setjmp.h>

static jmp_buf buf;

void second() {
    printf("second\n");         // prints
    longjmp(buf,1);             // jumps back to where setjmp was called - making setjmp now return 1
}

void first() {
    second();
    printf("first\n");          // does not print
}

int main() {   
    if (!setjmp(buf))
        first();                // when executed, setjmp returned 0
    else                        // when longjmp jumps back, setjmp returns 1
        printf("main\n");       // prints

    return 0;
}

When executed, the above program will output:

second
main

Notice that although the first() subroutine gets called, "first" is never printed. "main" gets printed as the conditional statement if (!setjmp(buf)) is executed a second time.

In this example, setjmp is used to bracket exception handling, like try in some other languages. The call to longjmp is analogous to a throw statement, allowing an exception to return an error status directly to the setjmp. The following code adheres to the 1999 ISO C standard and Single UNIX Specification by invoking setjmp in a limited range of contexts:[7]

• As the condition to an if, switch or iteration statement
• As above in conjunction with a single ! or comparison with an integer constant
• As a statement (with the return value unused)

Following these rules can make it easier for the implementation to create the environment buffer, which can be a sensitive operation.[3] More general use of setjmp can cause undefined behaviour, such as corruption of local variables; conforming compilers and environments are not required to protect or even warn against such usage. However, slightly more sophisticated idioms such as switch ((exception_type = setjmp(env))) { } are common in literature and practice, and remain relatively portable. A simple conforming methodology is presented below, where an additional variable is maintained along with the state buffer. This variable could be elaborated into a structure incorporating the buffer itself.

In a more modern-looking example, the usual "try" block would be implemented as a setjmp (with some preparation code for multilevel jumps, as seen in first), the "throw" as longjmp with the optional parameter as the exception, and the "catch" as the "else" block under "try".

#include <setjmp.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>

static void first();
static void second();

/* Use a file scoped static variable for the exception stack so we can access
 * it anywhere within this translation unit. */
static jmp_buf exception_env;
static int exception_type;

int main(void) {
    char* volatile mem_buffer = NULL;

    if (setjmp(exception_env)) {
        // if we get here there was an exception
        printf("first failed, exception type: %d\n", exception_type);
    } else {
        // Run code that may signal failure via longjmp.
        puts("calling first");
        first();

        mem_buffer = malloc(300); // allocate a resource
        printf("%s\n", strcpy(mem_buffer, "first succeeded")); // not reached
    }

    free(mem_buffer); // NULL can be passed to free, no operation is performed

    return 0;
}

static void first() {
    jmp_buf my_env;

    puts("entering first"); // reached

    memcpy(my_env, exception_env, sizeof my_env);

    switch (setjmp(exception_env)) {
        case 3: // if we get here there was an exception.
            puts("second failed, exception type: 3; remapping to type 1");
            exception_type = 1;

        default: // fall through
            memcpy(exception_env, my_env, sizeof exception_env); // restore exception stack
            longjmp(exception_env, exception_type); // continue handling the exception

        case 0: // normal, desired operation
            puts("calling second"); // reached 
            second();
            puts("second succeeded"); // not reached
    }

    memcpy(exception_env, my_env, sizeof exception_env); // restore exception stack

    puts("leaving first"); // never reached
}

static void second() {
    puts("entering second" ); // reached

    exception_type = 3;
    longjmp(exception_env, exception_type); // declare that the program has failed

    puts("leaving second"); // not reached
}

This program's output is:

calling first
entering first
calling second
entering second
second failed, exception type: 3; remapping to type 1
first failed, exception type: 1

Although the exception_type variable is technically not needed here since setjmp returns what nonzero value longjmp was called with (as in second and first), in practice a more elaborate global object would be used to accommodate richer exceptions.

In the real world, setjmp-longjmp (sjlj) used to be the standard way of exception handling in third-party Windows C++ compilers (namely MinGW), since the native Structured Exception Handling is poorly documented in general and was also patented on 32-bit Windows up until 2014.

C99 provides that longjmp is guaranteed to work only when the destination is a calling function, i.e., that the destination scope is guaranteed to be intact. Jumping to a function that has already terminated by return or longjmp is undefined.[6] However, most implementations of longjmp do not specifically destroy local variables when performing the jump. Since the context survives until its local variables are erased, it could actually be restored by setjmp. In many environments (such as Really Simple Threads and TinyTimbers), idioms such as if(!setjmp(child_env)) longjmp(caller_env); can allow a called function to effectively pause-and-resume at a setjmp.

This is exploited by thread libraries to provide cooperative multitasking facilities without using setcontext or other fiber facilities. Whereas setcontext is a library service which can create an execution context in heap-allocated memory and can support other services such as buffer overflow protection, abuse of setjmp is implemented by the programmer, who may reserve memory on the stack and fail to notify the library or operating system of the new operating context. On the other hand, a library's implementation of setcontext may internally use setjmp in a fashion similar to this example to save and restore a context, after it has been initialised somehow.

Considering that setjmp to a child function will generally work unless sabotaged, and setcontext, as part of POSIX, is not required to be provided by C implementations, this mechanism may be portable where the setcontext alternative fails.

Since no exception will be generated upon overflow of one of the multiple stacks in such a mechanism, it is essential to overestimate the space required for each context, including the one containing main() and including space for any signal handlers that might interrupt regular execution. Exceeding the allocated space will corrupt the other contexts, usually with the outermost functions first. Unfortunately, systems requiring this kind of programming strategy are often also small ones with limited resources.

#include <setjmp.h>
#include <stdio.h>

jmp_buf mainTask, childTask;

void call_with_cushion();
void child();

int main() {
    if (!setjmp(mainTask)) {
        call_with_cushion(); // child never returns, yield
    } // execution resumes after this "}" after first time that child yields

    while (1) {
        printf("Parent\n");
        
        if (!setjmp(mainTask))
            longjmp(childTask, 1); // yield - note that this is undefined under C99
    }
}

void call_with_cushion() {
    char space[1000]; // Reserve enough space for main to run
    space[999] = 1; // Do not optimize array out of existence
    child();
}

void child() {
    while (1) {
        printf("Child loop begin\n");
        
        if (!setjmp(childTask))
            longjmp(mainTask, 1); // yield - invalidates childTask in C99

        printf("Child loop end\n");

        if (!setjmp(childTask))
            longjmp(mainTask, 1); // yield - invalidates childTask in C99
    }

    /* Don't return. Instead we should set a flag to indicate that main()
       should stop yielding to us and then longjmp(mainTask, 1) */
}