ERR05-A. Application-independent code should provide error detection without dictating error handling

Application-independent code includes code that is:

• shipped with the compiler or operating system
• from a third-party library
• developed in-house

When application-specific code detects an error, it can respond on the spot with a specific action, as in:

if (something_really_bad_happens) {
  take_me_some_place_safe();
}

This is because the application must both detect errors and provide a mechanism for handling errors. But because application-independent code is not associated with any application, it cannot handle errors. However, it must still detect errors, and report them to an application, so that the application may handle them.

Error detection and reporting can take several forms:

• a return value (especially of type errno_t)
• an argument passed by address
• a global object (e.g., errno)
• longjmp()
• some combination of the above

Non-Compliant Code Example

This non-compliant code example consists of two application-independent functions f() and g(). The f() function is part of the external API for the module; the g() function is an internal function.

void g(void) {
  /* ... */
  if (something_really_bad_happens) {
    fprintf(stderr, "Something really bad happened!\n");
    abort();
  }
  /* ... */
}

void f(void) {
  g();
  /* ... do the rest of f ... */
}

If something_really_bad_happens in g(), the function prints an error message to stderr and then calls abort(). The problem is that this application-independent code does not know the context in which it is being called, so it is erroneous to handle the error.

[[Miller 04]] Practice 23 says:

When a library aborts due to some kind of anomaly, it is saying there is no hope for execution to proceed normally beyond the point where the anomaly is detected. Nonetheless, it is dictatorially making this decision on behalf of the client. Even if the anomaly turns out to be some kind of internal bug in the library, which obviously cannot be resolved in the current execution, aborting is a bad thing to do. The fact is, a library developer cannot possibly know the fault-tolerant context in which his/her library is being used. The client may indeed be able to recover from the situation even if the library cannot.

It is equally bad to eliminate the call to abort() from g(). In this case, there is no indication back to the calling function that any error has occurred.

Compliant Solution (Return Value)

One way to indicate errors is to return a value indicating success or errors. This compliant solution changes each function to return a value of type errno_t, where 0 indicates that no error has occurred.

const errno_t ESOMETHINGREALLYBAD = 1;

errno_t g(void) {
  /* ... */
  if (something_really_bad_happens) {
    return ESOMETHINGREALLYBAD;
  }
  /* ... */
  return 0;
}

errno_t f(void) {
  errno_t status;
  if ((status = g()) != 0)
    return status;
  /* ... do the rest of f ... */
  return 0;
}

A call to f() returns a status indicator which is zero upon success, and a non-zero value upon failure indicating what went wrong.

A return type of errno_t indicates that the function returns a status indicator (see DCL09-A. Declare functions that return an errno error code with a return type of errno_t).

While this solution is secure, it has the following drawbacks:

• Source and object code can significantly increase in size, perhaps by as much as 30-40% [[Saks 07b]]

• All function return values must be checked (see MEM32-C. Detect and handle memory allocation errors among many others.)
• Functions should not return other values if they return error indicators (see ERR02-A. Avoid in-band error indicators.)
• Any function that allocates resources must ensure they are freed incases where errors occur.

Compliant Solution (Address Argument)

Instead of encoding status indicators in the return value, each function can take a pointer as an argument, which is used to indicate errors. In the following example, each function uses a errno_t* argument to report errors.

const errno_t ESOMETHINGREALLYBAD = 1;

void g(errno_t* err) {
  if (err == NULL) {
    /* handle NULL pointer */
  }
  /* ... */
  if (something_really_bad_happens) {
    *err = ESOMETHINGREALLYBAD;
  } else {
    /* ... */
    *err = 0;
  }
}

void f(errno_t* err) {
  if (err == NULL) {
    /* handle NULL pointer */
  }
  g(err);
  if (*err == 0) {
    /* ... do the rest of f ... */
  }
  return 0;
}

A call to f() provides a status indicator that is zero upon success and a non-zero value upon failure, assuming the user provided a valid pointer to an object of type errno_t.

While this solution is secure, it has the following drawbacks:

• A return status can only be returned if the caller provides a valid pointer to an object of type errno_t. If this argument is NULL, there is no way to indicate this error.
• Source code becomes even larger due to the possibilities of receiving a null pointer.
• All error indicators must be checked after calling functions.
• Any function that allocates resources must ensure they are freed in cases where errors occur.
• Unlike return values, static analysis tools generally do not diagnose a failure to check error indicators passed as argument pointers.

Compliant Solution (Global Error Indicator)

Instead of encoding error indicators in the return value or arguments, a functions can indicate its status by assigning a value to a global variable. In the following example, each function uses a static indicator called my_errno.

The original errno variable was the Standard C library's implementation of error handling using this approach.

errno_t my_errno; /* also declared in a .h file */
const errno_t ESOMETHINGREALLYBAD = 1;

void g(void) {
  /* ... */
  if (something_really_bad_happens) {
    my_errno = ESOMETHINGREALLYBAD;
    return;
  }
  /* ... */
}

void f(void) {
  my_errno = 0;
  g();
  if (my_errno != 0) {
    return;
  }
  /* ... do the rest of f ... */
}

The call to f() provides a status indicator that is zero upon success and a non-zero value upon failure.

This solution has many of the same properties as those observed with errno, including advantages and drawbacks.

• Source code size is inflated, though not by as much as in other approaches.
• All error indicators must be checked after calling functions.
• Nesting of function calls that all use this mechanism is problematic.
• Any function that allocates resources must ensure they are freed in cases where errors occur.
• In general, combining registries of different sets of errors is difficult. For example, changing the above code to use errno is difficult and bug-prone because one must be precisely aware of when C library functions set and clear errno, and one must be aware of all valid errno values before adding new ones.
• There are major limitations on calling f() from other application-independent code. Because f() sets my_errno to 0, it may potentially be overwriting a nonzero error value set by another application-independent calling function.

For these reasons, among others, this approach is generally discouraged.

Compliant Solution ( setjmp() and longjmp() )

C provides two functions, setjmp() and longjmp(), that can be used to alter control flow. This allows a user of these functions to ignore error values and trust that control flow will be correctly diverted in the event of error.

The following example uses setjmp() and longjmp() to ensure that control flow is disrupted in the event of error, and also uses the my_errno indicator from the previous example.

#include <setjmp.h>

const errno_t ESOMETHINGREALLYBAD = 1;

jmp_buf exception_env;

void g(void) {
  /* ... */
  if (something_really_bad_happens) {
    longjmp(exception_env, ESOMETHINGREALLYBAD);
  }
  /* ... */
}

void f(void) {
  g();
  /* ... do the rest of f ... */
}

/* ... */
errno_t err;
if ((err = setjmp(exception_env)) != 0) {
  /* if we get here, an error occurred
     and err indicates what went wrong */
}
/* ... */
f();
/* if we get here, no errors occurred */
/* ... */

Calls to f() will either succeed or divert control into an if clause designed to catch the error.

• Source code will not become significantly larger because function signatures do not change, and neither do functions that neither detect nor handle the error.
• Allocated resources must still be freed despite the error.
• The application must call setjmp() before invoking application-independent code.
• Signals are not necessarily preserved through longjmp() calls.
• The use of setjmp()/longjmp() bypasses the normal function call and return discipline.
• Any function that allocates resources must ensure they are freed in cases where errors occur.

Summary

Risk Analysis

Lack of an error detection mechanism prevents applications from knowing when an error has disrupted normal program behavior.

Related Vulnerabilities

Search for vulnerabilities resulting from the violation of this rule on the CERT website.

References

[[Miller 04]]
[[Saks 07b]]

ERR04-A. Choose an appropriate termination strategy      12. Error Handling (ERR)       ERR06-A. Understand the termination behavior of assert() and abort()