Getting Started¶

#include "mutex.h"

class BankAccount {
private:
  Mutex mu;
  int   balance GUARDED_BY(mu);

  void depositImpl(int amount) {
    balance += amount;       // WARNING! Cannot write balance without locking mu.
  }

  void withdrawImpl(int amount) REQUIRES(mu) {
    balance -= amount;       // OK. Caller must have locked mu.
  }

public:
  void withdraw(int amount) {
    mu.Lock();
    withdrawImpl(amount);    // OK.  We've locked mu.
  }                          // WARNING!  Failed to unlock mu.

  void transferFrom(BankAccount& b, int amount) {
    mu.Lock();
    b.withdrawImpl(amount);  // WARNING!  Calling withdrawImpl() requires locking b.mu.
    depositImpl(amount);     // OK.  depositImpl() has no requirements.
    mu.Unlock();
  }
};

This example demonstrates the basic concepts behind the analysis. The GUARDED_BY attribute declares that a thread must lock mu before it can read or write to balance, thus ensuring that the increment and decrement operations are atomic. Similarly, REQUIRES declares that the calling thread must lock mu before calling withdrawImpl. Because the caller is assumed to have locked mu, it is safe to modify balance within the body of the method.

The depositImpl() method does not have REQUIRES, so the analysis issues a warning. Thread safety analysis is not inter-procedural, so caller requirements must be explicitly declared. There is also a warning in transferFrom(), because although the method locks this->mu, it does not lock b.mu. The analysis understands that these are two separate mutexes, in two different objects.

Finally, there is a warning in the withdraw() method, because it fails to unlock mu. Every lock must have a corresponding unlock, and the analysis will detect both double locks, and double unlocks. A function is allowed to acquire a lock without releasing it, (or vice versa), but it must be annotated as such (using ACQUIRE/RELEASE).

Running The Analysis¶

To run the analysis, simply compile with the -Wthread-safety flag, e.g.

clang -c -Wthread-safety example.cpp

Note that this example assumes the presence of a suitably annotated mutex.h that declares which methods perform locking, unlocking, and so on.

GUARDED_BY(c) and PT_GUARDED_BY(c)¶

GUARDED_BY is an attribute on data members, which declares that the data member is protected by the given capability. Read operations on the data require shared access, while write operations require exclusive access.

PT_GUARDED_BY is similar, but is intended for use on pointers and smart pointers. There is no constraint on the data member itself, but the data that it points to is protected by the given capability.

Mutex mu;
int *p1             GUARDED_BY(mu);
int *p2             PT_GUARDED_BY(mu);
unique_ptr<int> p3  PT_GUARDED_BY(mu);

void test() {
  p1 = 0;             // Warning!

  *p2 = 42;           // Warning!
  p2 = new int;       // OK.

  *p3 = 42;           // Warning!
  p3.reset(new int);  // OK.
}

REQUIRES(...), REQUIRES_SHARED(...)¶

Previously: EXCLUSIVE_LOCKS_REQUIRED, SHARED_LOCKS_REQUIRED

REQUIRES is an attribute on functions or methods, which declares that the calling thread must have exclusive access to the given capabilities. More than one capability may be specified. The capabilities must be held on entry to the function, and must still be held on exit.

REQUIRES_SHARED is similar, but requires only shared access.

Mutex mu1, mu2;
int a GUARDED_BY(mu1);
int b GUARDED_BY(mu2);

void foo() REQUIRES(mu1, mu2) {
  a = 0;
  b = 0;
}

void test() {
  mu1.Lock();
  foo();         // Warning!  Requires mu2.
  mu1.Unlock();
}

ACQUIRE(...), ACQUIRE_SHARED(...), RELEASE(...), RELEASE_SHARED(...)¶

Previously: EXCLUSIVE_LOCK_FUNCTION, SHARED_LOCK_FUNCTION, UNLOCK_FUNCTION

ACQUIRE is an attribute on functions or methods, which declares that the function acquires a capability, but does not release it. The caller must not hold the given capability on entry, and it will hold the capability on exit. ACQUIRE_SHARED is similar.

RELEASE and RELEASE_SHARED declare that the function releases the given capability. The caller must hold the capability on entry, and will no longer hold it on exit. It does not matter whether the given capability is shared or exclusive.

Mutex mu;
MyClass myObject GUARDED_BY(mu);

void lockAndInit() ACQUIRE(mu) {
  mu.Lock();
  myObject.init();
}

void cleanupAndUnlock() RELEASE(mu) {
  myObject.cleanup();
}                          // Warning!  Need to unlock mu.

void test() {
  lockAndInit();
  myObject.doSomething();
  cleanupAndUnlock();
  myObject.doSomething();  // Warning, mu is not locked.
}

If no argument is passed to ACQUIRE or RELEASE, then the argument is assumed to be this, and the analysis will not check the body of the function. This pattern is intended for use by classes which hide locking details behind an abstract interface. For example:

template <class T>
class CAPABILITY("mutex") Container {
private:
  Mutex mu;
  T* data;

public:
  // Hide mu from public interface.
  void Lock()   ACQUIRE() { mu.Lock(); }
  void Unlock() RELEASE() { mu.Unlock(); }

  T& getElem(int i) { return data[i]; }
};

void test() {
  Container<int> c;
  c.Lock();
  int i = c.getElem(0);
  c.Unlock();
}

EXCLUDES(...)¶

Previously: LOCKS_EXCLUDED

EXCLUDES is an attribute on functions or methods, which declares that the caller must not hold the given capabilities. This annotation is used to prevent deadlock. Many mutex implementations are not re-entrant, so deadlock can occur if the function acquires the mutex a second time.

Mutex mu;
int a GUARDED_BY(mu);

void clear() EXCLUDES(mu) {
  mu.Lock();
  a = 0;
  mu.Unlock();
}

void reset() {
  mu.Lock();
  clear();     // Warning!  Caller cannot hold 'mu'.
  mu.Unlock();
}

Unlike REQUIRES, EXCLUDES is optional. The analysis will not issue a warning if the attribute is missing, which can lead to false negatives in some cases. This issue is discussed further in Negative Capabilities.

NO_THREAD_SAFETY_ANALYSIS¶

NO_THREAD_SAFETY_ANALYSIS is an attribute on functions or methods, which turns off thread safety checking for that method. It provides an escape hatch for functions which are either (1) deliberately thread-unsafe, or (2) are thread-safe, but too complicated for the analysis to understand. Reasons for (2) will be described in the Known Limitations, below.

class Counter {
  Mutex mu;
  int a GUARDED_BY(mu);

  void unsafeIncrement() NO_THREAD_SAFETY_ANALYSIS { a++; }
};

Unlike the other attributes, NO_THREAD_SAFETY_ANALYSIS is not part of the interface of a function, and should thus be placed on the function definition (in the .cc or .cpp file) rather than on the function declaration (in the header).

RETURN_CAPABILITY(c)¶

Previously: LOCK_RETURNED

RETURN_CAPABILITY is an attribute on functions or methods, which declares that the function returns a reference to the given capability. It is used to annotate getter methods that return mutexes.

class MyClass {
private:
  Mutex mu;
  int a GUARDED_BY(mu);

public:
  Mutex* getMu() RETURN_CAPABILITY(mu) { return μ }

  // analysis knows that getMu() == mu
  void clear() REQUIRES(getMu()) { a = 0; }
};

ACQUIRED_BEFORE(...), ACQUIRED_AFTER(...)¶

ACQUIRED_BEFORE and ACQUIRED_AFTER are attributes on member declarations, specifically declarations of mutexes or other capabilities. These declarations enforce a particular order in which the mutexes must be acquired, in order to prevent deadlock.

Mutex m1;
Mutex m2 ACQUIRED_AFTER(m1);

// Alternative declaration
// Mutex m2;
// Mutex m1 ACQUIRED_BEFORE(m2);

void foo() {
  m2.Lock();
  m1.Lock();  // Warning!  m2 must be acquired after m1.
  m1.Unlock();
  m2.Unlock();
}

CAPABILITY(<string>)¶

Previously: LOCKABLE

CAPABILITY is an attribute on classes, which specifies that objects of the class can be used as a capability. The string argument specifies the kind of capability in error messages, e.g. "mutex". See the Container example given above, or the Mutex class in mutex.h.

SCOPED_CAPABILITY¶

Previously: SCOPED_LOCKABLE

SCOPED_CAPABILITY is an attribute on classes that implement RAII-style locking, in which a capability is acquired in the constructor, and released in the destructor. Such classes require special handling because the constructor and destructor refer to the capability via different names; see the MutexLocker class in mutex.h, below.

TRY_ACQUIRE(<bool>, ...), TRY_ACQUIRE_SHARED(<bool>, ...)¶

Previously: EXCLUSIVE_TRYLOCK_FUNCTION, SHARED_TRYLOCK_FUNCTION

These are attributes on a function or method that tries to acquire the given capability, and returns a boolean value indicating success or failure. The first argument must be true or false, to specify which return value indicates success, and the remaining arguments are interpreted in the same way as ACQUIRE. See mutex.h, below, for example uses.

ASSERT_CAPABILITY(...) and ASSERT_SHARED_CAPABILITY(...)¶

Previously: ASSERT_EXCLUSIVE_LOCK, ASSERT_SHARED_LOCK

These are attributes on a function or method that does a run-time test to see whether the calling thread holds the given capability. The function is assumed to fail (no return) if the capability is not held. See mutex.h, below, for example uses.

GUARDED_VAR and PT_GUARDED_VAR¶

Use of these attributes has been deprecated.

Warning flags¶

• -Wthread-safety: Umbrella flag which turns on the following three:

  • -Wthread-safety-attributes: Sanity checks on attribute syntax.

  • -Wthread-safety-analysis: The core analysis.

  • -Wthread-safety-precise: Requires that mutex expressions match precisely.

    This warning can be disabled for code which has a lot of aliases.

  • -Wthread-safety-reference: Checks when guarded members are passed by reference.

Negative Capabilities are an experimental feature, which are enabled with:

• -Wthread-safety-negative: Negative capabilities. Off by default.

When new features and checks are added to the analysis, they can often introduce additional warnings. Those warnings are initially released as beta warnings for a period of time, after which they are migrated into the standard analysis.

• -Wthread-safety-beta: New features. Off by default.

Lexical scope¶

Thread safety attributes contain ordinary C++ expressions, and thus follow ordinary C++ scoping rules. In particular, this means that mutexes and other capabilities must be declared before they can be used in an attribute. Use-before-declaration is okay within a single class, because attributes are parsed at the same time as method bodies. (C++ delays parsing of method bodies until the end of the class.) However, use-before-declaration is not allowed between classes, as illustrated below.

class Foo;

class Bar {
  void bar(Foo* f) REQUIRES(f->mu);  // Error: mu undeclared.
};

class Foo {
  Mutex mu;
};

Private Mutexes¶

Good software engineering practice dictates that mutexes should be private members, because the locking mechanism used by a thread-safe class is part of its internal implementation. However, private mutexes can sometimes leak into the public interface of a class. Thread safety attributes follow normal C++ access restrictions, so if mu is a private member of c, then it is an error to write c.mu in an attribute.

One workaround is to (ab)use the RETURN_CAPABILITY attribute to provide a public name for a private mutex, without actually exposing the underlying mutex. For example:

class MyClass {
private:
  Mutex mu;

public:
  // For thread safety analysis only.  Does not actually return mu.
  Mutex* getMu() RETURN_CAPABILITY(mu) { return 0; }

  void doSomething() REQUIRES(mu);
};

void doSomethingTwice(MyClass& c) REQUIRES(c.getMu()) {
  // The analysis thinks that c.getMu() == c.mu
  c.doSomething();
  c.doSomething();
}

In the above example, doSomethingTwice() is an external routine that requires c.mu to be locked, which cannot be declared directly because mu is private. This pattern is discouraged because it violates encapsulation, but it is sometimes necessary, especially when adding annotations to an existing code base. The workaround is to define getMu() as a fake getter method, which is provided only for the benefit of thread safety analysis.

No conditionally held locks.¶

The analysis must be able to determine whether a lock is held, or not held, at every program point. Thus, sections of code where a lock might be held will generate spurious warnings (false positives). For example:

void foo() {
  bool b = needsToLock();
  if (b) mu.Lock();
  ...  // Warning!  Mutex 'mu' is not held on every path through here.
  if (b) mu.Unlock();
}

No checking inside constructors and destructors.¶

The analysis currently does not do any checking inside constructors or destructors. In other words, every constructor and destructor is treated as if it was annotated with NO_THREAD_SAFETY_ANALYSIS. The reason for this is that during initialization, only one thread typically has access to the object which is being initialized, and it is thus safe (and common practice) to initialize guarded members without acquiring any locks. The same is true of destructors.

Ideally, the analysis would allow initialization of guarded members inside the object being initialized or destroyed, while still enforcing the usual access restrictions on everything else. However, this is difficult to enforce in practice, because in complex pointer-based data structures, it is hard to determine what data is owned by the enclosing object.

No inlining.¶

Thread safety analysis is strictly intra-procedural, just like ordinary type checking. It relies only on the declared attributes of a function, and will not attempt to inline any method calls. As a result, code such as the following will not work:

template<class T>
class AutoCleanup {
  T* object;
  void (T::*mp)();

public:
  AutoCleanup(T* obj, void (T::*imp)()) : object(obj), mp(imp) { }
  ~AutoCleanup() { (object->*mp)(); }
};

Mutex mu;
void foo() {
  mu.Lock();
  AutoCleanup<Mutex>(&mu, &Mutex::Unlock);
  // ...
}  // Warning, mu is not unlocked.

In this case, the destructor of Autocleanup calls mu.Unlock(), so the warning is bogus. However, thread safety analysis cannot see the unlock, because it does not attempt to inline the destructor. Moreover, there is no way to annotate the destructor, because the destructor is calling a function that is not statically known. This pattern is simply not supported.

No alias analysis.¶

The analysis currently does not track pointer aliases. Thus, there can be false positives if two pointers both point to the same mutex.

class MutexUnlocker {
  Mutex* mu;

public:
  MutexUnlocker(Mutex* m) RELEASE(m) : mu(m)  { mu->Unlock(); }
  ~MutexUnlocker() ACQUIRE(mu) { mu->Lock(); }
};

Mutex mutex;
void test() REQUIRES(mutex) {
  {
    MutexUnlocker munl(&mutex);  // unlocks mutex
    doSomeIO();
  }                              // Warning: locks munl.mu
}

The MutexUnlocker class is intended to be the dual of the MutexLocker class, defined in mutex.h. However, it doesn’t work because the analysis doesn’t know that munl.mu == mutex. The SCOPED_CAPABILITY attribute handles aliasing for MutexLocker, but does so only for that particular pattern.

ACQUIRED_BEFORE(...) and ACQUIRED_AFTER(...) are currently unimplemented.¶

To be fixed in a future update.