Compound operations are operations that consist of more than one discrete operation. Expressions that include postfix or prefix increment (++), postfix or prefix decrement (--), or compound assignment operators always result in compound operations. Compound assignment expressions use operators such as *=, /=, %=, +=, -=, <<=, >>=, ^=, and |=. Compound operations on shared variables must be performed atomically to prevent data races.

Noncompliant Code Example (Logical Negation)

This noncompliant code example declares a shared _Bool flag variable and provides a toggle_flag() method that negates the current value of flag:
#include <stdbool.h>
 
static bool flag = false;
 
void toggle_flag(void) {
  flag = !flag;
}
 
bool get_flag(void) {
  return flag;
}

Execution of this code may result in a data race because the value of flag is read, negated, and written back.

Consider, for example, two threads that call toggle_flag(). The expected effect of toggling flag twice is that it is restored to its original value. However, the following scenario leaves flag in the incorrect state:

Time

flag=

Thread

Action

1

true

t 1

Reads the current value of flag, true, into a cache

2

true

t 2

Reads the current value of flag, (still) true, into a different cache

3

true

t 1

Toggles the temporary variable in the cache to false

4

true

t 2

Toggles the temporary variable in the different cache to false

5

false

t 1

Writes the cache variable's value to flag

6

false

t 2

Writes the different cache variable's value to flag

As a result, the effect of the call by t 2 is not reflected in flag; the program behaves as if toggle_flag() was called only once, not twice.

Compliant Solution (Mutex)

This compliant solution restricts access to flag under a mutex lock:

#include <threads.h>
#include <stdbool.h>
 
static bool flag = false;
mtx_t flag_mutex;

/* Initialize flag_mutex */
bool init_mutex(int type) {
  /* Check mutex type */
  if (thrd_success != mtx_init(&flag_mutex, type)) {
    return false;  /* Report error */
  }
  return true;
}
 
void toggle_flag(void) {
  if (thrd_success != mtx_lock(&flag_mutex)) {
    /* Handle error */
  }
  flag = !flag;
  if (thrd_success != mtx_unlock(&flag_mutex)) {
    /* Handle error */
  }
}
 
bool get_flag(void) {
  bool temp_flag;
  if (thrd_success != mtx_lock(&flag_mutex)) {
    /* Handle error */
  }
  temp_flag = flag;
  if (thrd_success != mtx_unlock(&flag_mutex)) {
    /* Handle error */
  }
  return temp_flag;
}

This solution guards reads and writes to the flag field with a lock on the flag_mutex. This lock ensures that changes to flag are visible to all threads. Now, only two execution orders are possible. In one execution order, t1 obtains the mutex and completes the operation before t 2 can acquire the mutex, as shown here:

Time

flag=

Thread

Action

1

true

t 1

Reads the current value of flag, true, into a cache variable

2

true

t 1

Toggles the cache variable to false

3

false

t 1

Writes the cache variable's value to flag

4

false

t 2

Reads the current value of flag, false, into a different cache variable

5

false

t 2

Toggles the different cache variable to true

6

true

t 2

Writes the different cache variable's value to flag

The other execution order is similar, except that t 2 starts and finishes before t 1 

Compliant Solution (atomic_compare_exchange_weak() )

This compliant solution uses atomic variables and a compare-and-exchange operation to guarantee that the correct value is stored in flag. All updates are visible to other threads.

#include <stdatomic.h>
#include <stdbool.h>
 
static atomic_bool flag;

void init_flag(void) {
  atomic_init(&flag, false);
}
void toggle_flag(void) {
  bool old_flag = atomic_load(&flag);
  bool new_flag;
  do {
    new_flag = !old_flag;
  } while (!atomic_compare_exchange_weak(&flag, &old_flag, new_flag));
}
  
bool get_flag(void) {
  return atomic_load(&flag);
}

An alternative solution is to use the atomic_flag data type for managing Boolean values atomically.

Noncompliant Code Example (Addition of Primitives)

In this noncompliant code example, multiple threads can invoke the set_values() method to set the a and b fields. Because this code fails to test for integer overflow, users of this code must also ensure that the arguments to the set_values() method can be added without overflow (see INT32-C. Ensure that operations on signed integers do not result in overflow for more information).

static int a;
static int b;
 
int get_sum(void) {
  return a + b;
}
 
void set_values(int new_a, int new_b) {
  a = new_a;
  b = new_b;
}

The  get_sum() method contains a race condition. For example, when a and b currently have the values 0 and INT_MAX, respectively, and one thread calls get_sum() while another calls set_values(INT_MAX, 0), the get_sum() method might return either 0 or INT_MAX, or it might overflow. Overflow will occur when the first thread reads a and b after the second thread has set the value of a to INT_MAX but before it has set the value of b to 0.


Noncompliant Code Example (Addition of Atomic Integers)

In this noncompliant code example, a and b are replaced with atomic integers.


#include <stdatomic.h>

static atomic_int a;
static atomic_int b;

void init_ab(void) {
  atomic_init(&a, 0);
  atomic_init(&b, 0);
}

int get_sum(void) {
  return atomic_load(&a) + atomic_load(&b);
}
 
void set_values(int new_a, int new_b) {
  atomic_store(&a, new_a);
  atomic_store(&b, new_b);
}


The simple replacement of the two int fields with atomic integers fails to eliminate the race condition in the sum because the compound operation a.get() + b.get() is still non-atomic. While a sum of some value of a and some value of b will be returned, there is no guarantee that this value represents the sum of the values of a and b at any particular moment.

Compliant Solution (_Atomic struct)

This compliant solution uses an atomic struct, which guarantees that both numbers are read and written together.

#include <stdatomic.h>
  
static _Atomic struct ab_s {
  int a, b;
} ab;
 
void init_ab(void) {
  struct ab_s new_ab = {0, 0};
  atomic_init(&ab, new_ab);
}
 
int get_sum(void) {
  struct ab_s new_ab = atomic_load(&ab);
  return new_ab.a + new_ab.b;
}
 
void set_values(int new_a, int new_b) {
  struct ab_s new_ab = {new_a, new_b};
  atomic_store(&ab, new_ab);
}

On most modern platforms, this will compile to be lock-free.

Compliant Solution (Mutex)

This compliant solution protects the set_values() and get_sum() methods with a mutex to ensure atomicity:

#include <threads.h>
#include <stdbool.h>

static int a;
static int b;
mtx_t flag_mutex;

/* Initialize everything */
bool init_all(int type) {
  /* Check mutex type */
  a = 0;
  b = 0;
  if (thrd_success != mtx_init(&flag_mutex, type)) {
    return false;  /* Report error */
  }
  return true;
}
 
int get_sum(void) {
  if (thrd_success != mtx_lock(&flag_mutex)) {
    /* Handle error */
  }
  int sum = a + b;
  if (thrd_success != mtx_unlock(&flag_mutex)) {
    /* Handle error */
  }
  return sum;
}
 
void set_values(int new_a, int new_b) {
  if (thrd_success != mtx_lock(&flag_mutex)) {
    /* Handle error */
  }
  a = new_a;
  b = new_b;
  if (thrd_success != mtx_unlock(&flag_mutex)) {
    /* Handle error */
  }
}

Thanks to the mutex, it is now possible to add overflow checking to the get_sum()  function without introducing the possibility of a race condition.

Risk Assessment

When operations on shared variables are not atomic, unexpected results can be produced. For example, information can be disclosed inadvertently because one user can receive information about other users.

Rule

Severity

Likelihood

Remediation Cost

Priority

Level

CON07-C

Medium

Probable

Medium

P8

L2

Automated Detection

Tool

Version

Checker

Description

CodeSonar
8.1p0

CONCURRENCY.DATARACE

Data Race

Related Guidelines

CERT Oracle Secure Coding Standard for JavaVNA02-J. Ensure that compound operations on shared variables are atomic

MITRE CWE

CWE-366, Race condition within a thread
CWE-413, Improper resource locking
CWE-567, Unsynchronized access to shared data in a multithreaded context
CWE-667, Improper locking

Bibliography

[ISO/IEC 14882:2011]

Subclause 7.17, "Atomics"



   


7 Comments

  1. In the compliant solution for addition, why are we still using atomic_int now that all of the accesses are protected by a mutex?

    1. There are cases when you want both atomic variables and mutexes. But you are right; the compliant solution doesn't need its variables to be atomic, so I changed them back to non-atomic.

  2. I propose we add a second compliant solution for addition, that uses an atomic struct:

    #include <stdatomic.h>
     
    static _Atomic struct ab {
      int a, b;
    } ab;
    
    void init_ab(void) {
      struct ab new_ab = {0, 0};
      atomic_init(&ab, new_ab);
    }
    
    int get_sum(void) {
      struct ab new_ab = atomic_load(&ab);
      return new_ab.a + new_ab.b;
    }
    
    void set_values(int new_a, int new_b) {
      struct ab new_ab = {new_a, new_b};
      atomic_store(&ab, new_ab);
    }

    This has the advantage of being lock-free on most modern CPUs, and where locking is necessary, the compiler will do it for you.

    1. Unfortunately C17 has no concept of atomic structs.

      If you know of a platform that that code works on (say Windows), you could add this as a Windows-specific compliant solution.


      1. The standard says "The type modified by the _Atomic qualifier shall not be an array type or a function type", but doesn't say anything about structs. And GCC and Clang both compile my code fine with "-std=c17 -pedantic".

        1. Yow, ISO C does support atomic structs! There is *no* mention of them in s7.17 "Atomics", and it is only obliquely implied by 6.7.3p4 as you cited.

          So go ahead and add your code example as another compliant solution. Two things to tweak:
          * Don't call the struct and the static variable both 'ab'. Suggest calling the struct name 'ab_s' just to avoid confusion.
          * This solution also works under C11, so use that as the base standard.

          1. Thanks for the new code example. I tweaked the wording a bit. in particular, all these code examples are C11, so I took that out of the new example.