SEI CERT C Coding Standard

Space shortcuts

Page tree

A consistent locking policy guarantees that multiple threads cannot simultaneously access or modify shared data. Atomic variables eliminate the need for locks by guaranteeing thread safety when certain operations are performed on them. The thread-safe operations on atomic variables are specified in the C Standard, subclauses 7.17.7 and 7.17.8 [ISO/IEC 9899:2011]. While atomic operations can be combined, combined operations do not provide the thread safety provided by individual atomic operations.

Every time an atomic variable appears on the left side of an assignment operator, including a compound assignment operator such as *=, an atomic write is performed on the variable. The use of the increment (++) or decrement (--) operators on an atomic variable constitutes an atomic read-and-write operation and is consequently thread-safe. Any reference of an atomic variable anywhere else in an expression indicates a distinct atomic read on the variable.

If the same atomic variable appears twice in an expression, then two atomic reads, or an atomic read and an atomic write, are required. Such a pair of atomic operations is not thread-safe, as another thread can modify the atomic variable between the two operations. Consequently, an atomic variable must not be referenced twice in the same expression.

Noncompliant Code Example (atomic_bool)

This noncompliant code example declares a shared atomic_bool flag variable and provides a toggle_flag() method that negates the current value of flag:

#include <stdatomic.h>
#include <stdbool.h>
  
static atomic_bool flag = ATOMIC_VAR_INIT(false);
  
void init_flag(void) {
  atomic_init(&flag, false);
}
  
void toggle_flag(void) {
  bool temp_flag = atomic_load(&flag);
  temp_flag = !temp_flag;
  atomic_store(&flag, temp_flag);
}
    
bool get_flag(void) {
  return atomic_load(&flag);
}

Execution of this code may result in unexpected behavior because the value of flag is read, negated, and written back. This occurs even though the read and write are both atomic.

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 scenario in the following table leaves flag in the incorrect state.

toggle_flag() without Compare-and-Exchange

Time

flag

Thread

Action

1

true

t1

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

2

true

t2

Reads the current value of flag, which is still true, into a different cache

3

true

t1

Toggles the temporary variable in the cache to false

4

true

t2

Toggles the temporary variable in the different cache to false

5

false

t1

Writes the cache variable's value to flag

6

false

t2

Writes the different cache variable's value to flag

As a result, the effect of the call by t2 is not reflected in flag; the program behaves as if toggle_flag() was called only once, not twice.

Compliant Solution (atomic_compare_exchange_weak())

This compliant solution uses a compare-and-exchange to guarantee that the correct value is stored in flag. All updates are visible to other threads. The call to atomic_compare_exchange_weak() is in a loop in conformance with CON41-C. Wrap functions that can fail spuriously in a loop.

#include <stdatomic.h>
#include <stdbool.h>
 
static atomic_bool flag = ATOMIC_VAR_INIT(false);
 
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. However, atomic_flag does not support a toggle operation.

Compliant Solution (Compound Assignment)

This compliant solution uses the ^=  assignment operation to toggle flag. This operation is guaranteed to be atomic, according to the C Standard, 6.5.16.2, paragraph 3. This operation performs a bitwise-exclusive-or between its arguments, but for Boolean arguments, this is equivalent to negation.

#include <stdatomic.h>
#include <stdbool.h>
  
static atomic_bool flag = ATOMIC_VAR_INIT(false);
  
void toggle_flag(void) {
  flag ^= 1;
}
    
bool get_flag(void) {
  return flag;
}

An alternative solution is to use a mutex to protect the atomic operation, but this solution loses the performance benefits of atomic variables.

Noncompliant Code Example

This noncompliant code example takes an atomic global variable n and computes n + (n - 1) + (n - 2) + ... + 1, using the formula n * (n + 1) / 2:

#include <stdatomic.h>

atomic_int n = ATOMIC_VAR_INIT(0);
  
int compute_sum(void) {
  return n * (n + 1) / 2;
}

The value of n may change between the two atomic reads of n in the expression, yielding an incorrect result.

Compliant Solution

This compliant solution passes the atomic variable as a function argument, forcing the variable to be copied and guaranteeing a correct result. Note that the function's formal parameter need not be atomic, and the atomic variable can still be passed as an actual argument.

#include <stdatomic.h>
 
int compute_sum(int n) {
  return n * (n + 1) / 2;
}

Risk Assessment

When operations on atomic variables are assumed to be atomic, but are not atomic, surprising data races can occur, leading to corrupted data and invalid control flow.

Rule

Severity

Likelihood

Remediation Cost

Priority

Level

CON40-C

Medium

Probable

Medium

P8

L2

Automated Detection

Tool

Version

Checker

Description

Astrée
24.04
multiple-atomic-accessesPartially checked
Axivion Bauhaus Suite

7.2.0

CertC-CON40
CodeSonar
8.1p0

CONCURRENCY.MAA

Multiple Accesses of Atomic
Coverity
2017.07

EVALUATION_ORDER (partial)

MISRA 2012 Rule 13.2

VOLATILE_ATOICITY (possible)

Implemented
Helix QAC

2024.1

C1114, C1115, C1116

C++3171, C++4150


Klocwork
2024.1

CERT.CONC.ATOMIC_TWICE_EXPR


Parasoft C/C++test

2023.1

CERT_C-CON40-a

Do not refer to an atomic variable twice in an expression

Polyspace Bug Finder

R2023b

CERT C: Rule CON40-C

Checks for:

  • Atomic variable accessed twice in an expression
  • Atomic load and store sequence not atomic

Rule fully covered.

RuleChecker

24.04

multiple-atomic-accessesPartially checked

Related Vulnerabilities

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

Related Guidelines

Key here (explains table format and definitions)

Taxonomy

Taxonomy item

Relationship

CWE 2.11CWE-366, Race Condition within a Thread2017-07-07: CERT: Rule subset of CWE

CERT-CWE Mapping Notes

Key here for mapping notes

CWE-366 and CON40-C

CON40-C = Subset( CON43-C) Intersection( CON32-C, CON40-C) = Ø

CWE-366 = Union( CON40-C, list) where list =


  • C data races that do not involve an atomic variable used twice within an expression



Bibliography

[ISO/IEC 9899:2011]

6.5.16.2, "Compound Assignment"
7.17, "Atomics"


4 Comments

  1. Robert Seacord

    from Jürgen Schäfer:  The compliant solution

    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)); 
    }

    migt be changed to ...

    void toggle_flag(void) {
      bool new_flag;
      bool old_flag;
      do { 
        old_flag = atomic_load(&flag);
        new_flag = !old_flag; 
      } while (!atomic_compare_exchange_weak(&flag, &old_flag, new_flag)); 
    }

    The load/reload must be placed within the loop. Am I wrong? With kind regards Jürgen

  2. Robert Seacord

    No, the loop is simply because the atomic_compare_exchange_weak() function is permitted to fail spuriously, so we have to call it repeatedly until it succeeds.  The atomic_load() only needs to be called once, before the loop.

  3. Martin Sebor

    The following sentence below the first NCCE isn't correct:

    Execution of this code may result in a data race because the value of flag is read, negated, and written back. This occurs even though the read and write are both atomic.

    C defines a data race like this:

    The execution of a program contains a data race if it contains two conflicting actions in different threads, at least one of which is not atomic, and neither happens before the other.

    Since the flag variable is atomic and temp_flag is a local (i.e., distinct in each thread, and thus not subject to conflicting accesses), there is no data race.  (The program just may not have the intended effect.)

    1. Aaron Ballman

      I think you're correct, and so I've changed the wording to remove mention of data race.