CERT C++: CON43-C

Do not allow data races in multithreaded code

Description

This checker is deactivated in a default Polyspace^® as You Code analysis. See Checkers Deactivated in Polyspace as You Code Analysis (Polyspace Access).

Rule Definition

Do not allow data races in multithreaded code.¹

Polyspace Implementation

The rule checker checks for Data race.

Extend Checker

For the rule checker to detect issues, your code must use one of the concurrency primitives recognized by Polyspace for thread creation and shared variable protection. Otherwise, you must explicitly specify tasks, interrupts, critical sections, and other multitasking options in your project configuration. See Multitasking.

You can also extend this checker in the following ways:

• Map your multithreading functions to their known POSIX^® equivalents. See Extend Concurrency Defect Checkers to Unsupported Multithreading Environments.

• Detect all data races including ones involving atomic operations.

  Polyspace assumes that certain operations are atomic and excludes them from data race checks. See Define Atomic Operations in Multitasking Code. These assumptions might not apply to your environment. To extend the data race checkers to include these operations, use the option -detect-atomic-data-race. See Extend Data Race Checkers to Atomic Operations.

Examples

Data race

Issue

Data race occurs when:

• Multiple tasks perform unprotected operations on a shared variable.

• At least one task performs a write operation.

• At least one operation is nonatomic. To detect data race on both atomic and nonatomic operations, use the options -detect-atomic-data-race. See Extend Data Race Checkers to Atomic Operations.

  See Define Atomic Operations in Multitasking Code.

To find this defect, you must specify the multitasking options before analysis. To specify these options, on the Configuration pane, select Multitasking. For more information, see Configuring Polyspace Multitasking Analysis Manually.

Risk

Data race can result in unpredictable values of the shared variable because you do not control the order of the operations in different tasks.

Data races between two write operations are more serious than data races between a write and read operation. Two write operations can interfere with each other and result in indeterminate values. To identify write-write conflicts, use the filters on the Detail column of the Results List pane. For these conflicts, the Detail column shows the additional line:

 Variable value may be altered by write-write concurrent access.

See Filter and Group Results in Polyspace Desktop User Interface or Filter and Sort Results in Polyspace Access Web Interface (Polyspace Access).

Fix

To fix this defect, protect the operations on the shared variable using critical sections, temporal exclusion or another means. See Protections for Shared Variables in Multitasking Code.

To identify existing protections that you can reuse, see the table and graphs associated with the result. The table shows each pair of conflicting calls. The Access Protections column shows existing protections on the calls. To see the function call sequence leading to the conflicts, click the icon. For an example, see below.

Example - Unprotected Operation on Global Variable from Multiple Tasks

int var; //Noncompliant
void begin_critical_section(void);
void end_critical_section(void);

void increment(void) {
    var++; 
}

void task1(void)  { 
      increment();
}

void task2(void)  { 
      increment();
}

void task3(void)  { 
     begin_critical_section();
     increment();
     end_critical_section();
}

In this example, to emulate multitasking behavior, specify the following options:

Option	Specification
Configure multitasking manually
Tasks (-entry-points)	task1 task2 task3
Critical section details (-critical-section-begin -critical-section-end)	Starting routine	Ending routine
	begin_critical_section	end_critical_section

On the command-line, you can use the following:

 polyspace-bug-finder 
   -entry-points task1,task2,task3
   -critical-section-begin begin_critical_section:cs1
   -critical-section-end end_critical_section:cs1

In this example, the tasks task1, task2, and task3 call the function increment. increment contains the operation var++ that can involve multiple machine instructions including:

• Reading var.

• Writing an increased value to var.

These machine instructions, when executed from task1 and task2, can occur concurrently in an unpredictable sequence. For example, reading var from task1 can occur either before or after writing to var from task2. Therefore the value of var can be unpredictable.

Though task3 calls increment inside a critical section, other tasks do not use the same critical section. The operations in the critical section of task3 are not mutually exclusive with operations in other tasks.

Therefore, the three tasks are operating on a shared variable without common protection. In your result details, you see each pair of conflicting function calls.

If you click the icon, you see the function call sequence starting from the entry point to the read or write operation. You also see that the operation starting from task3 is in a critical section. The Access Protections entry shows the lock and unlock function that begin and end the critical section. In this example, you see the functions begin_critical_section and end_critical_section.

Correction — Place Operation in Critical Section

One possible correction is to place the operation in critical section. You can implement the critical section in multiple ways. For instance:

• You can place var++ in a critical section. When task1 enters its critical section, the other tasks cannot enter their critical sections until task1 leaves its critical section. The operation var++ from the three tasks cannot interfere with each other.

  To implement the critical section, in the function increment, place the operation var++ between calls to begin_critical_section and end_critical_section.

  
  
  int var;
  
  void begin_critical_section(void);
  void end_critical_section(void);
  
  void increment(void) {
        begin_critical_section();
        var++;
        end_critical_section(); 
  }
  
  void task1(void)  { 
        increment();
  }
  
  void task2(void)  { 
        increment();
  }
  
  void task3(void)  { 
        increment();
  }
  

• You can place the call to increment in the same critical section in the three tasks. When task1 enters its critical section, the other tasks cannot enter their critical sections until task1 leaves its critical section. The calls to increment from the three tasks cannot interfere with each other.

  To implement the critical section, in each of the three tasks, call increment between calls to begin_critical_section and end_critical_section.

  
  
  int var;
  
  void begin_critical_section(void);
  void end_critical_section(void);
  
  void increment(void) {
        var++;       
  }
  
  void task1(void)  { 
       begin_critical_section();
       increment();
       end_critical_section();
  }
  
  void task2(void)  { 
       begin_critical_section();
       increment();
       end_critical_section();
  }
  
  void task3(void)  { 
       begin_critical_section();
       increment();
       end_critical_section();
  }

Correction — Make Tasks Temporally Exclusive

Another possible correction is to make the tasks, task1, task2 and task3, temporally exclusive. Temporally exclusive tasks cannot execute concurrently.

On the Configuration pane, specify the following additional options:

Option	Value
Temporally exclusive tasks (-temporal-exclusions-file)	task1 task2 task3

On the command-line, you can use the following:

 polyspace-bug-finder 
     -temporal-exclusions-file "C:\exclusions_file.txt"

where the file C:\exclusions_file.txt has the following line:

task1 task2 task3

Example - Unprotected Operation in Threads Created with pthread_create

#include <pthread.h>

pthread_mutex_t count_mutex;
long long count; //Noncompliant


void* increment_count(void* args)
{
    count = count + 1;
    return NULL;
}

void* set_count(void *args)
{
    long long c;
    c = count;
    return NULL;
}

int main(void)
{
    pthread_t thread_increment;
    pthread_t thread_get;

    pthread_create(&thread_increment, NULL, increment_count, NULL);
    pthread_create(&thread_get, NULL, set_count, NULL);
    
    pthread_join(thread_get, NULL);
    pthread_join(thread_increment, NULL);

    return 1;
}

In this example, Bug Finder detects the creation of separate threads with pthread_create. The Data race defect is raised because the operation count = count + 1 in the thread with id thread_increment conflicts with the operation c = count in the thread with id thread_get. The variable count is accessed in multiple threads without a common protection.

The two conflicting operations are nonatomic. The operation c = count is nonatomic on 32-bit targets. See Define Atomic Operations in Multitasking Code.

Correction — Protect Operations with pthread_mutex_lock and pthread_mutex_unlock Pair

To prevent concurrent access on the variable count, protect operations on count with a critical section. Use the functions pthread_mutex_lock and pthread_mutex_unlock to implement the critical section.

#include <pthread.h>

pthread_mutex_t count_mutex;
long long count;


void* increment_count(void* args)
{
    pthread_mutex_lock(&count_mutex);
    count = count + 1;
    pthread_mutex_unlock(&count_mutex);
    return NULL;        
}

void* set_count(void *args)
{
    long long c;
    pthread_mutex_lock(&count_mutex);
    c = count;
    pthread_mutex_unlock(&count_mutex);
    return NULL;
}

int main(void)
{
    pthread_t thread_increment;
    pthread_t thread_get;

    pthread_create(&thread_increment, NULL, increment_count, NULL);
    pthread_create(&thread_get, NULL, set_count, NULL);

    pthread_join(thread_get, NULL);
    pthread_join(thread_increment, NULL);

    return 1;
}

Check Information

Group: 10. Concurrency (CON)

Version History

Introduced in R2019a

R2022a: New option -detect-atomic-data-race to detect data race involving possibly atomic operations

When checking for data races, the checker ignores operations that can be performed in one machine instruction. Previously, to check these atomic instructions for data races, you used -checkers with the argument data_race_all. This argument is no longer supported.

Starting in R2022a, use the option -detect-atomic-data-race to detect data races in operations that can be performed in one machine instruction.