// Copyright 2017 The Abseil Authors. // // Licensed under the Apache License, Version 2.0 (the "License"); // you may not use this file except in compliance with the License. // You may obtain a copy of the License at // // https://www.apache.org/licenses/LICENSE-2.0 // // Unless required by applicable law or agreed to in writing, software // distributed under the License is distributed on an "AS IS" BASIS, // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. // See the License for the specific language governing permissions and // limitations under the License. // // ----------------------------------------------------------------------------- // File: call_once.h // ----------------------------------------------------------------------------- // // This header file provides an Abseil version of `std::call_once` for invoking // a given function at most once, across all threads. This Abseil version is // faster than the C++11 version and incorporates the C++17 argument-passing // fix, so that (for example) non-const references may be passed to the invoked // function. #ifndef ABSL_BASE_CALL_ONCE_H_ #define ABSL_BASE_CALL_ONCE_H_ #include #include #include #include #include #include "absl/base/internal/invoke.h" #include "absl/base/internal/low_level_scheduling.h" #include "absl/base/internal/raw_logging.h" #include "absl/base/internal/scheduling_mode.h" #include "absl/base/internal/spinlock_wait.h" #include "absl/base/macros.h" #include "absl/base/optimization.h" #include "absl/base/port.h" namespace absl { class once_flag; namespace base_internal { std::atomic* ControlWord(absl::once_flag* flag); } // namespace base_internal // call_once() // // For all invocations using a given `once_flag`, invokes a given `fn` exactly // once across all threads. The first call to `call_once()` with a particular // `once_flag` argument (that does not throw an exception) will run the // specified function with the provided `args`; other calls with the same // `once_flag` argument will not run the function, but will wait // for the provided function to finish running (if it is still running). // // This mechanism provides a safe, simple, and fast mechanism for one-time // initialization in a multi-threaded process. // // Example: // // class MyInitClass { // public: // ... // mutable absl::once_flag once_; // // MyInitClass* init() const { // absl::call_once(once_, &MyInitClass::Init, this); // return ptr_; // } // template void call_once(absl::once_flag& flag, Callable&& fn, Args&&... args); // once_flag // // Objects of this type are used to distinguish calls to `call_once()` and // ensure the provided function is only invoked once across all threads. This // type is not copyable or movable. However, it has a `constexpr` // constructor, and is safe to use as a namespace-scoped global variable. class once_flag { public: constexpr once_flag() : control_(0) {} once_flag(const once_flag&) = delete; once_flag& operator=(const once_flag&) = delete; private: friend std::atomic* base_internal::ControlWord(once_flag* flag); std::atomic control_; }; //------------------------------------------------------------------------------ // End of public interfaces. // Implementation details follow. //------------------------------------------------------------------------------ namespace base_internal { // Like call_once, but uses KERNEL_ONLY scheduling. Intended to be used to // initialize entities used by the scheduler implementation. template void LowLevelCallOnce(absl::once_flag* flag, Callable&& fn, Args&&... args); // Disables scheduling while on stack when scheduling mode is non-cooperative. // No effect for cooperative scheduling modes. class SchedulingHelper { public: explicit SchedulingHelper(base_internal::SchedulingMode mode) : mode_(mode) { if (mode_ == base_internal::SCHEDULE_KERNEL_ONLY) { guard_result_ = base_internal::SchedulingGuard::DisableRescheduling(); } } ~SchedulingHelper() { if (mode_ == base_internal::SCHEDULE_KERNEL_ONLY) { base_internal::SchedulingGuard::EnableRescheduling(guard_result_); } } private: base_internal::SchedulingMode mode_; bool guard_result_; }; // Bit patterns for call_once state machine values. Internal implementation // detail, not for use by clients. // // The bit patterns are arbitrarily chosen from unlikely values, to aid in // debugging. However, kOnceInit must be 0, so that a zero-initialized // once_flag will be valid for immediate use. enum { kOnceInit = 0, kOnceRunning = 0x65C2937B, kOnceWaiter = 0x05A308D2, // A very small constant is chosen for kOnceDone so that it fit in a single // compare with immediate instruction for most common ISAs. This is verified // for x86, POWER and ARM. kOnceDone = 221, // Random Number }; template ABSL_ATTRIBUTE_NOINLINE void CallOnceImpl(std::atomic* control, base_internal::SchedulingMode scheduling_mode, Callable&& fn, Args&&... args) { #ifndef NDEBUG { uint32_t old_control = control->load(std::memory_order_relaxed); if (old_control != kOnceInit && old_control != kOnceRunning && old_control != kOnceWaiter && old_control != kOnceDone) { ABSL_RAW_LOG(FATAL, "Unexpected value for control word: 0x%lx", static_cast(old_control)); // NOLINT } } #endif // NDEBUG static const base_internal::SpinLockWaitTransition trans[] = { {kOnceInit, kOnceRunning, true}, {kOnceRunning, kOnceWaiter, false}, {kOnceDone, kOnceDone, true}}; // Must do this before potentially modifying control word's state. base_internal::SchedulingHelper maybe_disable_scheduling(scheduling_mode); // Short circuit the simplest case to avoid procedure call overhead. // The base_internal::SpinLockWait() call returns either kOnceInit or // kOnceDone. If it returns kOnceDone, it must have loaded the control word // with std::memory_order_acquire and seen a value of kOnceDone. uint32_t old_control = kOnceInit; if (control->compare_exchange_strong(old_control, kOnceRunning, std::memory_order_relaxed) || base_internal::SpinLockWait(control, ABSL_ARRAYSIZE(trans), trans, scheduling_mode) == kOnceInit) { base_internal::Invoke(std::forward(fn), std::forward(args)...); // The call to SpinLockWake below is an optimization, because the waiter // in SpinLockWait is waiting with a short timeout. The atomic load/store // sequence is slightly faster than an atomic exchange: // old_control = control->exchange(base_internal::kOnceDone, // std::memory_order_release); // We opt for a slightly faster case when there are no waiters, in spite // of longer tail latency when there are waiters. old_control = control->load(std::memory_order_relaxed); control->store(base_internal::kOnceDone, std::memory_order_release); if (old_control == base_internal::kOnceWaiter) { base_internal::SpinLockWake(control, true); } } // else *control is already kOnceDone } inline std::atomic* ControlWord(once_flag* flag) { return &flag->control_; } template void LowLevelCallOnce(absl::once_flag* flag, Callable&& fn, Args&&... args) { std::atomic* once = base_internal::ControlWord(flag); uint32_t s = once->load(std::memory_order_acquire); if (ABSL_PREDICT_FALSE(s != base_internal::kOnceDone)) { base_internal::CallOnceImpl(once, base_internal::SCHEDULE_KERNEL_ONLY, std::forward(fn), std::forward(args)...); } } } // namespace base_internal template void call_once(absl::once_flag& flag, Callable&& fn, Args&&... args) { std::atomic* once = base_internal::ControlWord(&flag); uint32_t s = once->load(std::memory_order_acquire); if (ABSL_PREDICT_FALSE(s != base_internal::kOnceDone)) { base_internal::CallOnceImpl( once, base_internal::SCHEDULE_COOPERATIVE_AND_KERNEL, std::forward(fn), std::forward(args)...); } } } // namespace absl #endif // ABSL_BASE_CALL_ONCE_H_