When multiple threads can read or modify the same data, use synchronization techniques to avoid software flaws that can lead to security vulnerabilities. Data races can often result in abnormal termination or denial of service, but it is possible for them to result in more serious vulnerabilities. The C Standard, section 126.96.36.199, paragraph 25 [ISO/IEC 9899:2011], says:
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. Any such data race results in undefined behavior.
Noncompliant Code Example
Assume this simplified code is part of a multithreaded bank system. Threads call
debit() as money is deposited into and withdrawn from the single account. Because the addition and subtraction operations are not atomic, it is possible that two operations can occur concurrently, but only the result of one would be saved—despite declaring the
volatile. For example, an attacker can credit the account with a sum of money and make a large number of small debits concurrently. Some of the debits might not affect the account balance because of the race condition, so the attacker is effectively creating money.
Compliant Solution (Mutex)
This compliant solution uses a mutex to make credits and debits atomic operations. All credits and debits will now affect the account balance, so an attacker cannot exploit the race condition to steal money from the bank. The mutex is created with the
mtx_init() function. The presence of the mutex makes declaring
Compliant Solution (Atomic)
This compliant solution uses an atomic variable to synchronize credit and debit operations. All credits and debits will now affect the account balance, so an attacker cannot exploit the race condition to steal money from the bank. The atomic integer does not need to be initialized because default (zero) initialization of an atomic object with static or thread-local storage is guaranteed to produce a valid state. The
-= operators behave atomically when used with an atomic variable.
Noncompliant Code Example (Double-Fetch)
This noncompliant code example illustrates Xen Security Advisory CVE-2015-8550 / XSA-155 In this example, the following code is vulnerable to a data race where the integer referenced by
ps could be modified by a second thread that ran between the two reads of the variable.
Even though there is only one read of the
*ps variable in the source code, the compiler is permitted to produce object code that performs multiple reads of the memory location. This is permitted by the "as-if" principle, as explained by section 5.1 of the [C99 Rationale 2003]:
The /as if/ principle is invoked repeatedly in this Rationale. The C89 Committee found that describing various aspects of the C language, library, and environment in terms of concrete models best serves discussion and presentation. Every attempt has been made to craft these models so that implementations are constrained only insofar as they must bring about the same result, /as if/ they had implemented the presentation model; often enough the clearest model would make for the worst implementation.
Implementation Details (GCC)
This code produces two reads of the
*ps value using GCC 4.8.4 on x86, as well as GCC 5.3.0 on x86-64 (Compiler-Introduced Double-Fetch Vulnerabilities – Understanding XSA-155).
Noncompliant Code Example (Volatile)
The data race can be disabled by declaring the data to be volatile, because the
volatile keyword forces the compiler to not produce two reads of the data. However, this violates CON02-C. Do not use volatile as a synchronization primitive.
Compliant Solution (C11, Atomic)
Declaring the data to be atomic also forces the compiler to produce only one read of the data.
Compliant Solution (C11, Fences)
The bug was actually resolved by erecting fences around the
Race conditions caused by multiple threads concurrently accessing and modifying the same data can lead to abnormal termination and denial-of-service attacks or data integrity violations.
|Supported by sound analysis (data race alarm)
Multiple Accesses of Atomic
C1765, C1766, C1770, C1771
C++1765, C++1766, C++1770, C++1771
Do not use global variable with different locks set
Partially supported: access is detected at the object level (not at the field level)
|Polyspace Bug Finder
|CERT C: Rule CON43-C
Checks for data race (rule fully covered)
Key here (explains table format and definitions)
|CWE-366, Race condition within a thread
|2017-07-07: CERT: Exact
|188.8.131.52, "Multi-threaded Executions and Data Races"
|[C99 Rationale 2003]
|Chapter 13, "Synchronization and State"
|Chapter 8, "File I/O"