Synchronization primitives for C++20: specialized mutex, semaphore and event types optimized for different use cases. Most notably this includes variants which can be integrated into native event loops (Boost.Asio, libdispatch, epoll, Qt, etc.), allowing to wait asynchronously for a lock or event without blocking the thread.

Futex-based mutex with configurable backoff.

• Policies:
  • with_backoff: Spins briefly with exponential backoff before parking, resulting in lower latency under brief contention.
  • timed: Enables timed waits via try_lock_for / try_lock_until.

Fair FIFO ticket lock guaranteeing strict acquisition order. Prevents starvation under sustained contention, but is not suitable for high-throughput low-contention workloads.

• Policies:
  • with_backoff: Spins with exponential backoff before parking.

Spinlock with different customization options.

• Policies:
  • with_backoff: Enables exponential backoff.
  • recursive: Allows re-entrant locking from the owning thread.
  • shared: Enables shared (reader-writer) locking via lock_shared() (mutually exclusive with recursive).

POSIX wrappers available when targeting POSIX systems.

• pthread_mutex: Wraps pthread_mutex_t.
  • Policies:
    • pthread_recursive: Enables re-entrant locking (PTHREAD_MUTEX_RECURSIVE).
    • pthread_errorcheck: Emits an error on double-lock (PTHREAD_MUTEX_ERRORCHECK).
    • pthread_adaptive: Enables adaptive-spin POSIX mutexes (Linux only, PTHREAD_MUTEX_ADAPTIVE_NP).
    • priority_inherit: Priority inheritance protocol where the owner is boosted to the highest waiter priority (PTHREAD_PRIO_INHERIT).
    • priority_ceiling<N>: Priority ceiling protocol where all holders are elevated to ceiling N (PTHREAD_PRIO_PROTECT).
• pthread_spinlock_mutex: Wraps pthread_spinlock_t.
• pthread_rwlock_mutex: Wraps pthread_rwlock_t for a POSIX reader-writer lock.

Note, priority_inherit and priority_ceiling policies typically require CAP_SYS_NICE or equivalent.

Mutexes provided by the Win32 runtime.

• win32_critical_section_mutex: Wraps a Windows CRITICAL_SECTION (which is recursive by default).
  • Policies: win32_spin_count<N> sets a custom spin count before falling back to a kernel wait.
• win32_srw_mutex: Wraps the Slim Reader/Writer (SRW) lock for an ultra-lightweight Windows mutex.

Mutexes provided by Apple's runtime.

• apple_os_unfair_mutex: Wraps os_unfair_lock.

win32_event_mutex (Windows), kqueue_mutex (macOS/iOS), and eventfd_mutex (Linux). These expose native OS handles (native_handle()) enabling integration with event loops (Boost.Asio, libdispatch, epoll, Qt, etc.).

• Policies:
  • with_backoff: Spins with exponential backoff before falling back to OS waits.
• Aliases:
  • native_async_mutex: Resolves to the pure async mutex for the current platform (e.g. kqueue_mutex<>).
  • native_fast_async_mutex: Resolves to the async mutex with backoff for the current platform (e.g. kqueue_mutex<with_backoff>).

Handlers receive an expected<std::unique_lock<Mutex>, std::error_code> (std::expected or tl::expected):

void handler(expected<std::unique_lock<Mutex>, std::error_code> result);

Boost.Asio — callback:

#include <nova/sync/mutex/boost_asio_support.hpp>

nova::sync::native_async_mutex mtx;
boost::asio::io_context        ioc;

// Non-cancellable:
nova::sync::async_acquire(ioc, mtx,
    [](auto result) {
        if (!result) return; // unexpected error
        auto& lock = *result; // lock.owns_lock() == true — critical section here
        // lock releases the mutex automatically on scope exit
    });

// Cancellable:
auto handle = nova::sync::async_acquire_cancellable(ioc, mtx,
    [](auto result) {
        if (!result) {
            if (result.error() == std::errc::operation_canceled) return; // cancelled
            return; // other error
        }
        auto& lock = *result; // lock.owns_lock() == true
    });
handle.cancel(); // abort the pending wait from any thread

Boost.Asio — future:

#include <nova/sync/mutex/boost_asio_support.hpp>

nova::sync::native_async_mutex mtx;
boost::asio::io_context        ioc;

auto [descriptor, fut] = nova::sync::async_acquire(ioc, mtx);
// descriptor keeps the wait alive; fut becomes ready when the lock is acquired

std::unique_lock lock = fut.get(); // blocks until acquired; lock.owns_lock() == true

// To cancel: descriptor->cancel(); // fut will never become ready

All mutex types are annotated for Clang's thread-safety analysis (-Wthread-safety). Macros in <nova/sync/thread_safety/macros.hpp> map to TSA attributes (e.g., NOVA_SYNC_GUARDED_BY, NOVA_SYNC_REQUIRES, NOVA_SYNC_EXCLUDES, NOVA_SYNC_ACQUIRE, NOVA_SYNC_RELEASE) on Clang and expand to nothing on other compilers.

Typical usage:

nova::sync::spinlock_mutex mtx;
int counter NOVA_SYNC_GUARDED_BY(mtx);

void increment() NOVA_SYNC_REQUIRES(mtx) { counter++; }

{
    nova::sync::lock_guard lock(mtx);
    increment();  // OK: mtx held by lock_guard
}
increment();      // Error: mutex not held

Type-safe RAII wrapper pairing a value T with a Mutex, enforcing synchronized via a smart-pointer style interface. The value is only accessible through lock guards. Supports exclusive locking (mutual exclusion) and shared locking (read-write patterns with std::shared_mutex or compatible).

#include <nova/sync/thread_safety/locked_object.hpp>

nova::sync::locked_object< int > counter( 0 );

// Exclusive lock on non-const instance → mutable access
{
    auto guard = counter.lock();  // returns locked_object_guard<int, std::mutex>
    *guard = 42;
}

// Const instance can acquire exclusive lock → const access
const auto& const_counter = counter;
{
    auto guard = const_counter.lock();  // returns locked_object_guard<const int, std::mutex>
    // *guard = 42;  // compile error: const access
    int val = *guard;  // OK
}

// Try-lock (non-blocking)
if (auto guard = counter.try_lock()) {
    *guard += 1;
}

// Shared lock (requires std::shared_mutex)
nova::sync::locked_object< std::vector<int>, std::shared_mutex > data( {} );
{
    auto guard = data.lock_shared();  // returns shared_locked_object_guard<...>
    for (int x : *guard) { /* ... */ }  // concurrent readers allowed
}

// Higher-order: lock_and (acquire lock, invoke function, auto-release)
counter.lock_and( [](int& v) {
    v = 100;
} );

int squared = counter.lock_and( [](const int& v) {
    return v * v;
} );

// Try-lock_and (non-blocking higher-order)
auto updated = counter.try_lock_and( [](int& v) {
    v += 10;
    return v;
} );  // returns std::optional<int>

// Shared lock higher-order
int sum = data.lock_shared_and( [](const std::vector<int>& v) {
    int result = 0;
    for (int x : v) result += x;
    return result;
} );

The following results were recorded on Ubuntu 25.04 on an Intel i7-14700K.

Single-threaded benchmark:

Multi-threaded benchmark:

Single-threaded benchmark:

Multi-threaded benchmark:

Single-threaded benchmark:

Multi-threaded benchmark:

Cross-platform lock-free counting semaphore based on futex.

• Policies:
  • with_backoff: Spins with exponential backoff before parking.

• timed_semaphore: Timed futex-based semaphore.
  • Policies: with_backoff enables exponential backoff.

These variants wrap OS primitives and expose native_handle() for integration with event loops (Boost.Asio, libdispatch, epoll, Qt, etc.).

• Windows: win32_semaphore.
• Linux: eventfd_semaphore.
• macOS / iOS: kqueue_semaphore.
• Aliases: native_async_semaphore resolves to the async semaphore with a native handle for the platform.

nova::sync::parking_semaphore sem(0);

// Producer thread
sem.release(5);  // add 5 tokens

// Consumer thread
sem.acquire();   // block until token available; consumes one
if (sem.try_acquire())  // non-blocking; consumes if available
    // ... token acquired

Async integration (with native_async_semaphore):

nova::sync::native_async_semaphore sem(0);

// Register for async notification when a token becomes available
auto handle = nova::sync::async_acquire_cancellable(ioc, sem,
    [](auto result) {
        if (result) {
            // Token acquired; use it
        } else if (result.error() == std::errc::operation_canceled) {
            // Wait was cancelled
        }
    });

handle.cancel();  // abort pending wait

• Linux: posix_semaphore (wrapper around sem_t).
• macOS / iOS: mach_semaphore (wrapper around semaphore_t) and dispatch_semaphore (wrapper around dispatch_semaphore_t).

Once signal() is called, all waiters are woken and subsequent wait() / try_wait() calls return immediately until reset() is called.

• parking_manual_reset_event: Futex-based manual reset event.
  • Policies: with_backoff enables exponential backoff.
• native_manual_reset_event: Maps to OS primitives (eventfd on Linux, kqueue on macOS, SetEvent on Windows) and exposes native_handle() for integration with event loops.

Each signal() delivers exactly one token. A blocked waiter consumes it; otherwise the next wait() / try_wait() call consumes it.

• parking_auto_reset_event: Futex-based auto reset event.
  • Policies: with_backoff enables exponential backoff.
• timed_auto_reset_event: Supports timed waits.
• native_auto_reset_event: Maps to OS primitives (eventfd on Linux, kqueue on macOS, SetEvent on Windows) and exposes native_handle() for integration with event loops.

nova::sync::parking_manual_reset_event ev;

// Producer thread
ev.signal();          // wake all waiters; event stays set

// Consumer threads
ev.wait();            // block until set
ev.try_wait();        // non-blocking check
ev.reset();           // clear the event

nova::sync::parking_auto_reset_event ev;

// Producer thread
ev.signal();          // deliver one token

// Consumer thread
ev.wait();            // block until a token is available; consumes it
ev.try_wait();        // non-blocking; returns true and consumes token if available

• C++20 (GCC 12+, Clang 17+, MSVC 2022+)
• No external dependencies for core library
• Tests require Catch2 and Boost.asio (fetched via CPM)

cmake -B build
cmake --build build
ctest --test-dir build

MIT — see License.txt