Redaguoti

Bendrinti naudojant


Annotating Locking Behavior

To avoid concurrency bugs in your multithreaded program, always follow an appropriate locking discipline and use SAL annotations.

Concurrency bugs are notoriously hard to reproduce, diagnose, and debug because they're nondeterministic. Reasoning about thread interleaving is difficult at best, and becomes impractical when you're designing a body of code that has more than a few threads. Therefore, it's good practice to follow a locking discipline in your multithreaded programs. For example, obeying a lock order while acquiring multiple locks helps avoid deadlocks, and acquiring the proper guarding lock before accessing a shared resource helps prevent race conditions.

Unfortunately, seemingly simple locking rules can be surprisingly hard to follow in practice. A fundamental limitation in today's programming languages and compilers is that they don't directly support the specification and analysis of concurrency requirements. Programmers have to rely on informal code comments to express their intentions about how they use locks.

Concurrency SAL annotations are designed to help you specify locking side effects, locking responsibility, data guardianship, lock order hierarchy, and other expected locking behavior. By making implicit rules explicit, SAL concurrency annotations provide a consistent way for you to document how your code uses locking rules. Concurrency annotations also enhance the ability of code analysis tools to find race conditions, deadlocks, mismatched synchronization operations, and other subtle concurrency errors.

General Guidelines

By using annotations, you can state the contracts that are implied by function definitions between implementations (callees) and clients (callers). You can also express invariants and other properties of the program that can further improve analysis.

SAL supports many different kinds of locking primitives—for example, critical sections, mutexes, spin locks, and other resource objects. Many concurrency annotations take a lock expression as a parameter. By convention, a lock is denoted by the path expression of the underlying lock object.

Some thread ownership rules to keep in mind:

  • Spin locks are uncounted locks that have clear thread ownership.

  • Mutexes and critical sections are counted locks that have clear thread ownership.

  • Semaphores and events are counted locks that don't have clear thread ownership.

Locking Annotations

The following table lists the locking annotations.

Annotation Description
_Acquires_exclusive_lock_(expr) Annotates a function and indicates that in post state the function increments by one the exclusive lock count of the lock object that's named by expr.
_Acquires_lock_(expr) Annotates a function and indicates that in post state the function increments by one the lock count of the lock object that's named by expr.
_Acquires_nonreentrant_lock_(expr) The lock that's named by expr is acquired. An error is reported if the lock is already held.
_Acquires_shared_lock_(expr) Annotates a function and indicates that in post state the function increments by one the shared lock count of the lock object that's named by expr.
_Create_lock_level_(name) A statement that declares the symbol name to be a lock level so that it may be used in the annotations _Has_Lock_level_ and _Lock_level_order_.
_Has_lock_kind_(kind) Annotates any object to refine the type information of a resource object. Sometimes a common type is used for different kinds of resources and the overloaded type isn't sufficient to distinguish the semantic requirements among various resources. Here's a list of predefined kind parameters:

_Lock_kind_mutex_
Lock kind ID for mutexes.

_Lock_kind_event_
Lock kind ID for events.

_Lock_kind_semaphore_
Lock kind ID for semaphores.

_Lock_kind_spin_lock_
Lock kind ID for spin locks.

_Lock_kind_critical_section_
Lock kind ID for critical sections.
_Has_lock_level_(name) Annotates a lock object and gives it the lock level of name.
_Lock_level_order_(name1, name2) A statement that gives the lock ordering between name1 and name2. Locks that have level name1 must be acquired before locks that have level name2.
_Post_same_lock_(dst, src) Annotates a function and indicates that in post state the two locks, dst and src, are treated as if they're the same lock object, by applying lock properties from src to dst.
_Releases_exclusive_lock_(expr) Annotates a function and indicates that in post state the function decrements by one the exclusive lock count of the lock object that's named by expr.
_Releases_lock_(expr) Annotates a function and indicates that in post state the function decrements by one the lock count of the lock object that's named by expr.
_Releases_nonreentrant_lock_(expr) The lock that's named by expr is released. An error is reported if the lock isn't currently held.
_Releases_shared_lock_(expr) Annotates a function and indicates that in post state the function decrements by one the shared lock count of the lock object that's named by expr.
_Requires_lock_held_(expr) Annotates a function and indicates that in pre state the lock count of the object that's named by expr is at least one.
_Requires_lock_not_held_(expr) Annotates a function and indicates that in pre state the lock count of the object that's named by expr is zero.
_Requires_no_locks_held_ Annotates a function and indicates that the lock counts of all locks that are known to the checker are zero.
_Requires_shared_lock_held_(expr) Annotates a function and indicates that in pre state the shared lock count of the object that's named by expr is at least one.
_Requires_exclusive_lock_held_(expr) Annotates a function and indicates that in pre state the exclusive lock count of the object that's named by expr is at least one.

SAL Intrinsics For Unexposed Locking Objects

Certain lock objects aren't exposed by the implementation of the associated locking functions. The following table lists SAL intrinsic variables that enable annotations on functions that operate on those unexposed lock objects.

Annotation Description
_Global_cancel_spin_lock_ Describes the cancel spin lock.
_Global_critical_region_ Describes the critical region.
_Global_interlock_ Describes interlocked operations.
_Global_priority_region_ Describes the priority region.

Shared Data Access Annotations

The following table lists the annotations for shared data access.

Annotation Description
_Guarded_by_(expr) Annotates a variable and indicates that whenever the variable is accessed, the lock count of the lock object that's named by expr is at least one.
_Interlocked_ Annotates a variable and is equivalent to _Guarded_by_(_Global_interlock_).
_Interlocked_operand_ The annotated function parameter is the target operand of one of the various Interlocked functions. Those operands must have other specific properties.
_Write_guarded_by_(expr) Annotates a variable and indicates that whenever the variable is modified, the lock count of the lock object that's named by expr is at least one.

Smart Lock and RAII Annotations

Smart locks typically wrap native locks and manage their lifetime. The following table lists annotations that can be used with smart locks and Resource Acquisition Is Initialization (RAII) coding patterns with support for move semantics.

Annotation Description
_Analysis_assume_smart_lock_acquired_(lock) Tells the analyzer to assume that a smart lock was acquired. This annotation expects a reference lock type as its parameter.
_Analysis_assume_smart_lock_released_(lock) Tells the analyzer to assume that a smart lock was released. This annotation expects a reference lock type as its parameter.
_Moves_lock_(target, source) Describes a move constructor operation, which transfers lock state from the source object to the target. The target is considered a newly constructed object, so any state it had before is lost and replaced by the source state. The source is also reset to a clean state with no lock counts or aliasing target, but aliases pointing to it remain unchanged.
_Replaces_lock_(target, source) Describes move assignment operator semantics where the target lock is released before transferring the state from the source. You can regard it as a combination of _Moves_lock_(target, source) preceded by a _Releases_lock_(target).
_Swaps_locks_(left, right) Describes the standard swap behavior, which assumes that objects left and right exchange their state. The state exchanged includes lock count and aliasing target, if present. Aliases that point to the left and right objects remain unchanged.
_Detaches_lock_(detached, lock) Describes a scenario in which a lock wrapper type allows dissociation with its contained resource. It's similar to how std::unique_ptr works with its internal pointer: it allows programmers to extract the pointer and leave its smart pointer container in a clean state. Similar logic is supported by std::unique_lock and can be implemented in custom lock wrappers. The detached lock retains its state (lock count and aliasing target, if any), while the wrapper is reset to contain zero lock count and no aliasing target, while retaining its own aliases. There's no operation on lock counts (releasing and acquiring). This annotation behaves exactly as _Moves_lock_ except that the detached argument should be return rather than this.

See also