The setjmp() macro should be invoked from only one of the contexts listed in subclause 7.13.1.1 of the C Standard [ISO/IEC 9899:2011]. Invoking setjmp() outside of one of these contexts results in undefined behavior. (See undefined behavior 125.)

After invoking longjmp(), non-volatile-qualified local objects should not be accessed if their values could have changed since the invocation of setjmp(). Their value in this case is considered indeterminate, and accessing them is undefined behavior. (See undefined behaviors 127 and 10.)

The longjmp() function should never be used to return control to a function that has terminated execution. (See undefined behavior 126.)

Signal masks, floating-point status flags, and the state of open files are not saved by the setjmp() function. If signal masks need to be saved, the POSIX sigsetjmp() function should be used.

This recommendation is related to SIG30-C. Call only asynchronous-safe functions within signal handlers and ENV32-C. All exit handlers must return normally.

Noncompliant Code Example

This noncompliant code example calls setjmp() in an assignment statement, resulting in undefined behavior:

jmp_buf buf;

void f(void) {
  int i = setjmp(buf);
  if (i == 0) {
    g();
  } else {
    /* longjmp was invoked */
  }
}

void g(void) {
  /* ... */
  longjmp(buf, 1);
}

Compliant Solution

Placing the call to setjmp() in the if statement and, optionally, comparing it with a constant integer removes the undefined behavior, as shown in this compliant solution:

jmp_buf buf;

void f(void) {
  if (setjmp(buf) == 0) {
    g();
  } else {
    /* longjmp was invoked */
  }
}

void g(void) {
  /* ... */
  longjmp(buf, 1);
}

Noncompliant Code Example

Any attempt to invoke the longjmp() function to transfer control to a function that has completed execution results in undefined behavior:

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

static jmp_buf buf;
static void bad(void);

static void g(void) {
  if (setjmp(buf) == 0) {
    printf("setjmp() invoked\n");
  } else {
    printf("longjmp() invoked\n");
  }
}

static void f(void) {
  g();
}

static void setup(void) {
  f();
}

void do_stuff(void) {
  void (*b)(void) = bad;
  /* ... */
  longjmp(buf, 1);
}

static void bad(void) {
  printf("Should not be called!\n");
  exit(1);
}

int main(void) {
  setup();
  do_stuff();
}

Implementation Details

Compiled at -O0 using GCC 7.5 or Clang 8.0 on Ubuntu 18.04 (Linux for x86-64), the preceding example outputs the following when run:

setjmp() invoked
longjmp() invoked
Should not be called!

Because g() has finished executing at the time longjmp() is called, it is no longer on the stack. When do_stuff() is invoked, its stack frame occupies the same memory as the old stack frame of g(). In this case, a was located in the same location as the return address of function g(). The assignment of b overwrites the return address, so when longjmp() sends control back to function g(), the function returns to the wrong address (in this case, to function bad()).

If the array b were user specified, the user would be able to set the return address of function g() to any location.

Compliant Solution

The longjmp() function should be used only when the function containing the corresponding setjmp() is guaranteed not to have completed execution, as in the following example:

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

static jmp_buf buf;
static void bad(void);

void do_stuff(void) {
  void (*b)(void) = bad;
  /* ... */
  longjmp(buf, 1);
}

static void bad(void) {
  printf("Should not be called!\n");
  exit(1);
}

int main(void) {
  if (setjmp(buf) == 0) {
    printf("setjmp() invoked\n");
  } else {
    printf("longjmp() invoked\n");
  }
  do_stuff();
} 

There is no risk of overwriting a return address because the stack frame of main() (the function that invoked setjmp()) is still on the stack; so when do_stuff() is invoked, the two stack frames will not overlap.

Noncompliant Code Example

In this noncompliant example, non-volatile-qualified objects local to the function that invoked the corresponding setjmp() have indeterminate values after longjmp() is executed if their value has been changed since the invocation of setjmp():

jmp_buf buf;

void f(void) {
  int i = 0;
  if (setjmp(buf) != 0) {
    printf("%i\n", i);
    /* ... */
  }
  i = 2;
  g();
}

void g(void) {
  /* ... */
  longjmp(buf, 1);
}

Implementation Details

Calling f() will print 2 if you compile with -O0, but will print 0 if you compile with -O2.  This involves using GCC 7.5 or Clang 8.0 on Ubuntu 18.04 (Linux x86-64).

Compliant Solution

If an object local to the function that invoked setjmp() needs to be accessed after longjmp() returns control to the function, the object should be volatile-qualified:

jmp_buf buf;

void f(void) {
  volatile int i = 0;
  if (setjmp(buf) != 0) {
    printf("%i\n", i);
    /* ... */
  }
  i = 2;
  g();
}

void g(void) {
  /* ... */
  longjmp(buf, 1);
}

This will now correctly print 2 regardless of optimization level.

Risk Assessment

Recommendation

Severity

Likelihood

Remediation Cost

Priority

Level

MSC22-C

Low

Probable

Medium

P4

L3

Automated Detection

ToolVersionCheckerDescription
CodeSonar
8.1p0

BADFUNC.LONGJMP

BADFUNC.SETJMP

Use of longjmp

Use of setjmp

LDRA tool suite
9.7.1
43 SEnhanced enforcement
Parasoft C/C++test
2023.1

CERT_C-MSC22-a

The facilities provided by <setjmp.h> should not be used
Polyspace Bug Finder

R2023b

CERT C: Rec. MSC22-CChecks for use of setjmp/longjmp (rec. fully covered)


SonarQube C/C++ Plugin
3.11
S982



9 Comments

  1. Good rule so far. Comments:

    • I disagree with the Remediation Cost
    • The two references can be merged since they come from the same document

    Any attempt to longjmp to a function that has terminated execution results in possibly exploitable undefined behavior.

    • How is it exploitable? Please provide details or remove that phrase. Perhaps an 'Implementation Details' section describing what actually happens when you run the NCCE on some platform would be helpful.
    • By our style guidelines, you should say longjmp() rather than just longjmp.
    1. I agree that Remediation Cost was off; I've changed it to medium. Two of the three rules should be automatically detectable (using setjmp() in the wrong place, and accessing non-volatile local objects), though I'm not sure if you can automatically detect a longjmp() to a terminated functions. If you can't, then should the remediation cost be set to high, or is medium good?

      I've taken out the 'possibly exploitable' phrase for now. If I can confirm that is exploitable, I'll add in an example.

      1. I'm willing to bet that (unless you are doing weird things with function pointers), you can always tell if a longjmp() call might occur at some point in the program where a setjmp() might not be in the call stack. That's a 99% guarantee, which IMHO is good enough to justify a medium remediation cost.

        Good rule so far, but I'm really hungry for implementation details on various platforms.

  2. In the second compliant solution, the behavior of the program is undefined when h() is invoked the second time, after the if statement in f() terminates (in other words, the if statement must return from f() in order for the solution to be compliant).

    jmp_buf buf;
    
    void f() {
      if (setjmp(buf) != 0) {
        /* longjmp was invoked */
      }
      h();   // called twice, once when setjmp(buf) == 0 holds and again when it doesn't
      return;
    }
    
    void h() {
      /* ... */
      longjmp(buf, 1);
    }
    
    1. Is this actually undefined behavior? I can't find anything in the C99 standard saying multiple longjmp() invocations to the same setjmp() is undefined behavior.

      Regardless, h() wasn't supposed to be called twice. It's fixed. Thanks for pointing it out

      1. You're right, it's perfectly fine. There's even an example in the spec that looks similar to your original. I must have been thinking of va_list.

    • The 'Implementation Details' after the 2nd CS is unnecessary, as no undefined behavior occurs. If you like, the text there can just appear after the CS code itself.
  3. The implementation details concerning the glibc v2.11.1 definition of the jmp_buf type seems like a distraction. Does this add anything to this guideline, or can it be removed?

  4. Now that I think of it, shouldn't this be a rule? These calls are easy to misuse, and doing so is always UB. It may not be easy for a tool to detect misuse, but given that they are rare (used at most once in any program that uses them at all), not much onus on an auditor.