From 4f847581b1f31cc171d6fbdcea9a3c605566a89d Mon Sep 17 00:00:00 2001 From: Francesco Montorsi Date: Tue, 13 Aug 2019 11:19:36 +0200 Subject: [PATCH 01/11] Implement a very simple zero-lock message pool to test performance gains by completely removing malloc/free from the hot path of benchmark util --- include/atomicops.h | 676 +++++++++++++++++++++++++++ include/readerwriterqueue.h | 906 ++++++++++++++++++++++++++++++++++++ perf/remote_thr.cpp | 108 +++++ src/msg.cpp | 28 +- 4 files changed, 1717 insertions(+), 1 deletion(-) create mode 100644 include/atomicops.h create mode 100644 include/readerwriterqueue.h diff --git a/include/atomicops.h b/include/atomicops.h new file mode 100644 index 0000000000..4fd1748293 --- /dev/null +++ b/include/atomicops.h @@ -0,0 +1,676 @@ +// ©2013-2016 Cameron Desrochers. +// Distributed under the simplified BSD license (see the license file that +// should have come with this header). +// Uses Jeff Preshing's semaphore implementation (under the terms of its +// separate zlib license, embedded below). + +#pragma once + +// Provides portable (VC++2010+, Intel ICC 13, GCC 4.7+, and anything C++11 compliant) implementation +// of low-level memory barriers, plus a few semi-portable utility macros (for inlining and alignment). +// Also has a basic atomic type (limited to hardware-supported atomics with no memory ordering guarantees). +// Uses the AE_* prefix for macros (historical reasons), and the "moodycamel" namespace for symbols. + +#include +#include +#include +#include +#include + +// Platform detection +#if defined(__INTEL_COMPILER) +#define AE_ICC +#elif defined(_MSC_VER) +#define AE_VCPP +#elif defined(__GNUC__) +#define AE_GCC +#endif + +#if defined(_M_IA64) || defined(__ia64__) +#define AE_ARCH_IA64 +#elif defined(_WIN64) || defined(__amd64__) || defined(_M_X64) || defined(__x86_64__) +#define AE_ARCH_X64 +#elif defined(_M_IX86) || defined(__i386__) +#define AE_ARCH_X86 +#elif defined(_M_PPC) || defined(__powerpc__) +#define AE_ARCH_PPC +#else +#define AE_ARCH_UNKNOWN +#endif + + +// AE_UNUSED +#define AE_UNUSED(x) ((void)x) + +// AE_NO_TSAN +#if defined(__has_feature) +#if __has_feature(thread_sanitizer) +#define AE_NO_TSAN __attribute__((no_sanitize("thread"))) +#else +#define AE_NO_TSAN +#endif +#else +#define AE_NO_TSAN +#endif + + +// AE_FORCEINLINE +#if defined(AE_VCPP) || defined(AE_ICC) +#define AE_FORCEINLINE __forceinline +#elif defined(AE_GCC) +//#define AE_FORCEINLINE __attribute__((always_inline)) +#define AE_FORCEINLINE inline +#else +#define AE_FORCEINLINE inline +#endif + + +// AE_ALIGN +#if defined(AE_VCPP) || defined(AE_ICC) +#define AE_ALIGN(x) __declspec(align(x)) +#elif defined(AE_GCC) +#define AE_ALIGN(x) __attribute__((aligned(x))) +#else +// Assume GCC compliant syntax... +#define AE_ALIGN(x) __attribute__((aligned(x))) +#endif + + +// Portable atomic fences implemented below: + +namespace moodycamel { + +enum memory_order { + memory_order_relaxed, + memory_order_acquire, + memory_order_release, + memory_order_acq_rel, + memory_order_seq_cst, + + // memory_order_sync: Forces a full sync: + // #LoadLoad, #LoadStore, #StoreStore, and most significantly, #StoreLoad + memory_order_sync = memory_order_seq_cst +}; + +} // end namespace moodycamel + +#if (defined(AE_VCPP) && (_MSC_VER < 1700 || defined(__cplusplus_cli))) || (defined(AE_ICC) && __INTEL_COMPILER < 1600) +// VS2010 and ICC13 don't support std::atomic_*_fence, implement our own fences + +#include + +#if defined(AE_ARCH_X64) || defined(AE_ARCH_X86) +#define AeFullSync _mm_mfence +#define AeLiteSync _mm_mfence +#elif defined(AE_ARCH_IA64) +#define AeFullSync __mf +#define AeLiteSync __mf +#elif defined(AE_ARCH_PPC) +#include +#define AeFullSync __sync +#define AeLiteSync __lwsync +#endif + + +#ifdef AE_VCPP +#pragma warning(push) +#pragma warning(disable: 4365) // Disable erroneous 'conversion from long to unsigned int, signed/unsigned mismatch' error when using `assert` +#ifdef __cplusplus_cli +#pragma managed(push, off) +#endif +#endif + +namespace moodycamel { + +AE_FORCEINLINE void compiler_fence(memory_order order) AE_NO_TSAN +{ + switch (order) { + case memory_order_relaxed: break; + case memory_order_acquire: _ReadBarrier(); break; + case memory_order_release: _WriteBarrier(); break; + case memory_order_acq_rel: _ReadWriteBarrier(); break; + case memory_order_seq_cst: _ReadWriteBarrier(); break; + default: assert(false); + } +} + +// x86/x64 have a strong memory model -- all loads and stores have +// acquire and release semantics automatically (so only need compiler +// barriers for those). +#if defined(AE_ARCH_X86) || defined(AE_ARCH_X64) +AE_FORCEINLINE void fence(memory_order order) AE_NO_TSAN +{ + switch (order) { + case memory_order_relaxed: break; + case memory_order_acquire: _ReadBarrier(); break; + case memory_order_release: _WriteBarrier(); break; + case memory_order_acq_rel: _ReadWriteBarrier(); break; + case memory_order_seq_cst: + _ReadWriteBarrier(); + AeFullSync(); + _ReadWriteBarrier(); + break; + default: assert(false); + } +} +#else +AE_FORCEINLINE void fence(memory_order order) AE_NO_TSAN +{ + // Non-specialized arch, use heavier memory barriers everywhere just in case :-( + switch (order) { + case memory_order_relaxed: + break; + case memory_order_acquire: + _ReadBarrier(); + AeLiteSync(); + _ReadBarrier(); + break; + case memory_order_release: + _WriteBarrier(); + AeLiteSync(); + _WriteBarrier(); + break; + case memory_order_acq_rel: + _ReadWriteBarrier(); + AeLiteSync(); + _ReadWriteBarrier(); + break; + case memory_order_seq_cst: + _ReadWriteBarrier(); + AeFullSync(); + _ReadWriteBarrier(); + break; + default: assert(false); + } +} +#endif +} // end namespace moodycamel +#else +// Use standard library of atomics +#include + +namespace moodycamel { + +AE_FORCEINLINE void compiler_fence(memory_order order) AE_NO_TSAN +{ + switch (order) { + case memory_order_relaxed: break; + case memory_order_acquire: std::atomic_signal_fence(std::memory_order_acquire); break; + case memory_order_release: std::atomic_signal_fence(std::memory_order_release); break; + case memory_order_acq_rel: std::atomic_signal_fence(std::memory_order_acq_rel); break; + case memory_order_seq_cst: std::atomic_signal_fence(std::memory_order_seq_cst); break; + default: assert(false); + } +} + +AE_FORCEINLINE void fence(memory_order order) AE_NO_TSAN +{ + switch (order) { + case memory_order_relaxed: break; + case memory_order_acquire: std::atomic_thread_fence(std::memory_order_acquire); break; + case memory_order_release: std::atomic_thread_fence(std::memory_order_release); break; + case memory_order_acq_rel: std::atomic_thread_fence(std::memory_order_acq_rel); break; + case memory_order_seq_cst: std::atomic_thread_fence(std::memory_order_seq_cst); break; + default: assert(false); + } +} + +} // end namespace moodycamel + +#endif + + +#if !defined(AE_VCPP) || (_MSC_VER >= 1700 && !defined(__cplusplus_cli)) +#define AE_USE_STD_ATOMIC_FOR_WEAK_ATOMIC +#endif + +#ifdef AE_USE_STD_ATOMIC_FOR_WEAK_ATOMIC +#include +#endif +#include + +// WARNING: *NOT* A REPLACEMENT FOR std::atomic. READ CAREFULLY: +// Provides basic support for atomic variables -- no memory ordering guarantees are provided. +// The guarantee of atomicity is only made for types that already have atomic load and store guarantees +// at the hardware level -- on most platforms this generally means aligned pointers and integers (only). +namespace moodycamel { +template +class weak_atomic +{ +public: + AE_NO_TSAN weak_atomic() { } +#ifdef AE_VCPP +#pragma warning(push) +#pragma warning(disable: 4100) // Get rid of (erroneous) 'unreferenced formal parameter' warning +#endif + template AE_NO_TSAN weak_atomic(U&& x) : value(std::forward(x)) { } +#ifdef __cplusplus_cli + // Work around bug with universal reference/nullptr combination that only appears when /clr is on + AE_NO_TSAN weak_atomic(nullptr_t) : value(nullptr) { } +#endif + AE_NO_TSAN weak_atomic(weak_atomic const& other) : value(other.load()) { } + AE_NO_TSAN weak_atomic(weak_atomic&& other) : value(std::move(other.load())) { } +#ifdef AE_VCPP +#pragma warning(pop) +#endif + + AE_FORCEINLINE operator T() const AE_NO_TSAN { return load(); } + + +#ifndef AE_USE_STD_ATOMIC_FOR_WEAK_ATOMIC + template AE_FORCEINLINE weak_atomic const& operator=(U&& x) AE_NO_TSAN { value = std::forward(x); return *this; } + AE_FORCEINLINE weak_atomic const& operator=(weak_atomic const& other) AE_NO_TSAN { value = other.value; return *this; } + + AE_FORCEINLINE T load() const AE_NO_TSAN { return value; } + + AE_FORCEINLINE T fetch_add_acquire(T increment) AE_NO_TSAN + { +#if defined(AE_ARCH_X64) || defined(AE_ARCH_X86) + if (sizeof(T) == 4) return _InterlockedExchangeAdd((long volatile*)&value, (long)increment); +#if defined(_M_AMD64) + else if (sizeof(T) == 8) return _InterlockedExchangeAdd64((long long volatile*)&value, (long long)increment); +#endif +#else +#error Unsupported platform +#endif + assert(false && "T must be either a 32 or 64 bit type"); + return value; + } + + AE_FORCEINLINE T fetch_add_release(T increment) AE_NO_TSAN + { +#if defined(AE_ARCH_X64) || defined(AE_ARCH_X86) + if (sizeof(T) == 4) return _InterlockedExchangeAdd((long volatile*)&value, (long)increment); +#if defined(_M_AMD64) + else if (sizeof(T) == 8) return _InterlockedExchangeAdd64((long long volatile*)&value, (long long)increment); +#endif +#else +#error Unsupported platform +#endif + assert(false && "T must be either a 32 or 64 bit type"); + return value; + } +#else + template + AE_FORCEINLINE weak_atomic const& operator=(U&& x) AE_NO_TSAN + { + value.store(std::forward(x), std::memory_order_relaxed); + return *this; + } + + AE_FORCEINLINE weak_atomic const& operator=(weak_atomic const& other) AE_NO_TSAN + { + value.store(other.value.load(std::memory_order_relaxed), std::memory_order_relaxed); + return *this; + } + + AE_FORCEINLINE T load() const AE_NO_TSAN { return value.load(std::memory_order_relaxed); } + + AE_FORCEINLINE T fetch_add_acquire(T increment) AE_NO_TSAN + { + return value.fetch_add(increment, std::memory_order_acquire); + } + + AE_FORCEINLINE T fetch_add_release(T increment) AE_NO_TSAN + { + return value.fetch_add(increment, std::memory_order_release); + } +#endif + + +private: +#ifndef AE_USE_STD_ATOMIC_FOR_WEAK_ATOMIC + // No std::atomic support, but still need to circumvent compiler optimizations. + // `volatile` will make memory access slow, but is guaranteed to be reliable. + volatile T value; +#else + std::atomic value; +#endif +}; + +} // end namespace moodycamel + + + +// Portable single-producer, single-consumer semaphore below: + +#if defined(_WIN32) +// Avoid including windows.h in a header; we only need a handful of +// items, so we'll redeclare them here (this is relatively safe since +// the API generally has to remain stable between Windows versions). +// I know this is an ugly hack but it still beats polluting the global +// namespace with thousands of generic names or adding a .cpp for nothing. +extern "C" { + struct _SECURITY_ATTRIBUTES; + __declspec(dllimport) void* __stdcall CreateSemaphoreW(_SECURITY_ATTRIBUTES* lpSemaphoreAttributes, long lInitialCount, long lMaximumCount, const wchar_t* lpName); + __declspec(dllimport) int __stdcall CloseHandle(void* hObject); + __declspec(dllimport) unsigned long __stdcall WaitForSingleObject(void* hHandle, unsigned long dwMilliseconds); + __declspec(dllimport) int __stdcall ReleaseSemaphore(void* hSemaphore, long lReleaseCount, long* lpPreviousCount); +} +#elif defined(__MACH__) +#include +#elif defined(__unix__) +#include +#endif + +namespace moodycamel +{ + // Code in the spsc_sema namespace below is an adaptation of Jeff Preshing's + // portable + lightweight semaphore implementations, originally from + // https://github.com/preshing/cpp11-on-multicore/blob/master/common/sema.h + // LICENSE: + // Copyright (c) 2015 Jeff Preshing + // + // This software is provided 'as-is', without any express or implied + // warranty. In no event will the authors be held liable for any damages + // arising from the use of this software. + // + // Permission is granted to anyone to use this software for any purpose, + // including commercial applications, and to alter it and redistribute it + // freely, subject to the following restrictions: + // + // 1. The origin of this software must not be misrepresented; you must not + // claim that you wrote the original software. If you use this software + // in a product, an acknowledgement in the product documentation would be + // appreciated but is not required. + // 2. Altered source versions must be plainly marked as such, and must not be + // misrepresented as being the original software. + // 3. This notice may not be removed or altered from any source distribution. + namespace spsc_sema + { +#if defined(_WIN32) + class Semaphore + { + private: + void* m_hSema; + + Semaphore(const Semaphore& other); + Semaphore& operator=(const Semaphore& other); + + public: + AE_NO_TSAN Semaphore(int initialCount = 0) + { + assert(initialCount >= 0); + const long maxLong = 0x7fffffff; + m_hSema = CreateSemaphoreW(nullptr, initialCount, maxLong, nullptr); + } + + AE_NO_TSAN ~Semaphore() + { + CloseHandle(m_hSema); + } + + void wait() AE_NO_TSAN + { + const unsigned long infinite = 0xffffffff; + WaitForSingleObject(m_hSema, infinite); + } + + bool try_wait() AE_NO_TSAN + { + const unsigned long RC_WAIT_TIMEOUT = 0x00000102; + return WaitForSingleObject(m_hSema, 0) != RC_WAIT_TIMEOUT; + } + + bool timed_wait(std::uint64_t usecs) AE_NO_TSAN + { + const unsigned long RC_WAIT_TIMEOUT = 0x00000102; + return WaitForSingleObject(m_hSema, (unsigned long)(usecs / 1000)) != RC_WAIT_TIMEOUT; + } + + void signal(int count = 1) AE_NO_TSAN + { + ReleaseSemaphore(m_hSema, count, nullptr); + } + }; +#elif defined(__MACH__) + //--------------------------------------------------------- + // Semaphore (Apple iOS and OSX) + // Can't use POSIX semaphores due to http://lists.apple.com/archives/darwin-kernel/2009/Apr/msg00010.html + //--------------------------------------------------------- + class Semaphore + { + private: + semaphore_t m_sema; + + Semaphore(const Semaphore& other); + Semaphore& operator=(const Semaphore& other); + + public: + AE_NO_TSAN Semaphore(int initialCount = 0) + { + assert(initialCount >= 0); + semaphore_create(mach_task_self(), &m_sema, SYNC_POLICY_FIFO, initialCount); + } + + AE_NO_TSAN ~Semaphore() + { + semaphore_destroy(mach_task_self(), m_sema); + } + + void wait() AE_NO_TSAN + { + semaphore_wait(m_sema); + } + + bool try_wait() AE_NO_TSAN + { + return timed_wait(0); + } + + bool timed_wait(std::int64_t timeout_usecs) AE_NO_TSAN + { + mach_timespec_t ts; + ts.tv_sec = static_cast(timeout_usecs / 1000000); + ts.tv_nsec = (timeout_usecs % 1000000) * 1000; + + // added in OSX 10.10: https://developer.apple.com/library/prerelease/mac/documentation/General/Reference/APIDiffsMacOSX10_10SeedDiff/modules/Darwin.html + kern_return_t rc = semaphore_timedwait(m_sema, ts); + + return rc != KERN_OPERATION_TIMED_OUT && rc != KERN_ABORTED; + } + + void signal() AE_NO_TSAN + { + semaphore_signal(m_sema); + } + + void signal(int count) AE_NO_TSAN + { + while (count-- > 0) + { + semaphore_signal(m_sema); + } + } + }; +#elif defined(__unix__) + //--------------------------------------------------------- + // Semaphore (POSIX, Linux) + //--------------------------------------------------------- + class Semaphore + { + private: + sem_t m_sema; + + Semaphore(const Semaphore& other); + Semaphore& operator=(const Semaphore& other); + + public: + AE_NO_TSAN Semaphore(int initialCount = 0) + { + assert(initialCount >= 0); + sem_init(&m_sema, 0, initialCount); + } + + AE_NO_TSAN ~Semaphore() + { + sem_destroy(&m_sema); + } + + void wait() AE_NO_TSAN + { + // http://stackoverflow.com/questions/2013181/gdb-causes-sem-wait-to-fail-with-eintr-error + int rc; + do + { + rc = sem_wait(&m_sema); + } + while (rc == -1 && errno == EINTR); + } + + bool try_wait() AE_NO_TSAN + { + int rc; + do { + rc = sem_trywait(&m_sema); + } while (rc == -1 && errno == EINTR); + return !(rc == -1 && errno == EAGAIN); + } + + bool timed_wait(std::uint64_t usecs) AE_NO_TSAN + { + struct timespec ts; + const int usecs_in_1_sec = 1000000; + const int nsecs_in_1_sec = 1000000000; + clock_gettime(CLOCK_REALTIME, &ts); + ts.tv_sec += usecs / usecs_in_1_sec; + ts.tv_nsec += (usecs % usecs_in_1_sec) * 1000; + // sem_timedwait bombs if you have more than 1e9 in tv_nsec + // so we have to clean things up before passing it in + if (ts.tv_nsec >= nsecs_in_1_sec) { + ts.tv_nsec -= nsecs_in_1_sec; + ++ts.tv_sec; + } + + int rc; + do { + rc = sem_timedwait(&m_sema, &ts); + } while (rc == -1 && errno == EINTR); + return !(rc == -1 && errno == ETIMEDOUT); + } + + void signal() AE_NO_TSAN + { + sem_post(&m_sema); + } + + void signal(int count) AE_NO_TSAN + { + while (count-- > 0) + { + sem_post(&m_sema); + } + } + }; +#else +#error Unsupported platform! (No semaphore wrapper available) +#endif + + //--------------------------------------------------------- + // LightweightSemaphore + //--------------------------------------------------------- + class LightweightSemaphore + { + public: + typedef std::make_signed::type ssize_t; + + private: + weak_atomic m_count; + Semaphore m_sema; + + bool waitWithPartialSpinning(std::int64_t timeout_usecs = -1) AE_NO_TSAN + { + ssize_t oldCount; + // Is there a better way to set the initial spin count? + // If we lower it to 1000, testBenaphore becomes 15x slower on my Core i7-5930K Windows PC, + // as threads start hitting the kernel semaphore. + int spin = 10000; + while (--spin >= 0) + { + if (m_count.load() > 0) + { + m_count.fetch_add_acquire(-1); + return true; + } + compiler_fence(memory_order_acquire); // Prevent the compiler from collapsing the loop. + } + oldCount = m_count.fetch_add_acquire(-1); + if (oldCount > 0) + return true; + if (timeout_usecs < 0) + { + m_sema.wait(); + return true; + } + if (m_sema.timed_wait(timeout_usecs)) + return true; + // At this point, we've timed out waiting for the semaphore, but the + // count is still decremented indicating we may still be waiting on + // it. So we have to re-adjust the count, but only if the semaphore + // wasn't signaled enough times for us too since then. If it was, we + // need to release the semaphore too. + while (true) + { + oldCount = m_count.fetch_add_release(1); + if (oldCount < 0) + return false; // successfully restored things to the way they were + // Oh, the producer thread just signaled the semaphore after all. Try again: + oldCount = m_count.fetch_add_acquire(-1); + if (oldCount > 0 && m_sema.try_wait()) + return true; + } + } + + public: + AE_NO_TSAN LightweightSemaphore(ssize_t initialCount = 0) : m_count(initialCount) + { + assert(initialCount >= 0); + } + + bool tryWait() AE_NO_TSAN + { + if (m_count.load() > 0) + { + m_count.fetch_add_acquire(-1); + return true; + } + return false; + } + + void wait() AE_NO_TSAN + { + if (!tryWait()) + waitWithPartialSpinning(); + } + + bool wait(std::int64_t timeout_usecs) AE_NO_TSAN + { + return tryWait() || waitWithPartialSpinning(timeout_usecs); + } + + void signal(ssize_t count = 1) AE_NO_TSAN + { + assert(count >= 0); + ssize_t oldCount = m_count.fetch_add_release(count); + assert(oldCount >= -1); + if (oldCount < 0) + { + m_sema.signal(1); + } + } + + ssize_t availableApprox() const AE_NO_TSAN + { + ssize_t count = m_count.load(); + return count > 0 ? count : 0; + } + }; + } // end namespace spsc_sema +} // end namespace moodycamel + +#if defined(AE_VCPP) && (_MSC_VER < 1700 || defined(__cplusplus_cli)) +#pragma warning(pop) +#ifdef __cplusplus_cli +#pragma managed(pop) +#endif +#endif diff --git a/include/readerwriterqueue.h b/include/readerwriterqueue.h new file mode 100644 index 0000000000..071147c3e1 --- /dev/null +++ b/include/readerwriterqueue.h @@ -0,0 +1,906 @@ +// ©2013-2016 Cameron Desrochers. +// Distributed under the simplified BSD license (see the license file that +// should have come with this header). + +#pragma once + +#include "atomicops.h" +#include +#include +#include +#include +#include +#include +#include // For malloc/free/abort & size_t +#include +#if __cplusplus > 199711L || _MSC_VER >= 1700 // C++11 or VS2012 +#include +#endif + + +// A lock-free queue for a single-consumer, single-producer architecture. +// The queue is also wait-free in the common path (except if more memory +// needs to be allocated, in which case malloc is called). +// Allocates memory sparingly (O(lg(n) times, amortized), and only once if +// the original maximum size estimate is never exceeded. +// Tested on x86/x64 processors, but semantics should be correct for all +// architectures (given the right implementations in atomicops.h), provided +// that aligned integer and pointer accesses are naturally atomic. +// Note that there should only be one consumer thread and producer thread; +// Switching roles of the threads, or using multiple consecutive threads for +// one role, is not safe unless properly synchronized. +// Using the queue exclusively from one thread is fine, though a bit silly. + +#ifndef MOODYCAMEL_CACHE_LINE_SIZE +#define MOODYCAMEL_CACHE_LINE_SIZE 64 +#endif + +#ifndef MOODYCAMEL_EXCEPTIONS_ENABLED +#if (defined(_MSC_VER) && defined(_CPPUNWIND)) || (defined(__GNUC__) && defined(__EXCEPTIONS)) || (!defined(_MSC_VER) && !defined(__GNUC__)) +#define MOODYCAMEL_EXCEPTIONS_ENABLED +#endif +#endif + +#ifndef MOODYCAMEL_HAS_EMPLACE +#if !defined(_MSC_VER) || _MSC_VER >= 1800 // variadic templates: either a non-MS compiler or VS >= 2013 +#define MOODYCAMEL_HAS_EMPLACE 1 +#endif +#endif + +#ifdef AE_VCPP +#pragma warning(push) +#pragma warning(disable: 4324) // structure was padded due to __declspec(align()) +#pragma warning(disable: 4820) // padding was added +#pragma warning(disable: 4127) // conditional expression is constant +#endif + +namespace moodycamel { + +template +class ReaderWriterQueue +{ + // Design: Based on a queue-of-queues. The low-level queues are just + // circular buffers with front and tail indices indicating where the + // next element to dequeue is and where the next element can be enqueued, + // respectively. Each low-level queue is called a "block". Each block + // wastes exactly one element's worth of space to keep the design simple + // (if front == tail then the queue is empty, and can't be full). + // The high-level queue is a circular linked list of blocks; again there + // is a front and tail, but this time they are pointers to the blocks. + // The front block is where the next element to be dequeued is, provided + // the block is not empty. The back block is where elements are to be + // enqueued, provided the block is not full. + // The producer thread owns all the tail indices/pointers. The consumer + // thread owns all the front indices/pointers. Both threads read each + // other's variables, but only the owning thread updates them. E.g. After + // the consumer reads the producer's tail, the tail may change before the + // consumer is done dequeuing an object, but the consumer knows the tail + // will never go backwards, only forwards. + // If there is no room to enqueue an object, an additional block (of + // equal size to the last block) is added. Blocks are never removed. + +public: + typedef T value_type; + + // Constructs a queue that can hold maxSize elements without further + // allocations. If more than MAX_BLOCK_SIZE elements are requested, + // then several blocks of MAX_BLOCK_SIZE each are reserved (including + // at least one extra buffer block). + AE_NO_TSAN explicit ReaderWriterQueue(size_t maxSize = 15) +#ifndef NDEBUG + : enqueuing(false) + ,dequeuing(false) +#endif + { + assert(maxSize > 0); + assert(MAX_BLOCK_SIZE == ceilToPow2(MAX_BLOCK_SIZE) && "MAX_BLOCK_SIZE must be a power of 2"); + assert(MAX_BLOCK_SIZE >= 2 && "MAX_BLOCK_SIZE must be at least 2"); + + Block* firstBlock = nullptr; + + largestBlockSize = ceilToPow2(maxSize + 1); // We need a spare slot to fit maxSize elements in the block + if (largestBlockSize > MAX_BLOCK_SIZE * 2) { + // We need a spare block in case the producer is writing to a different block the consumer is reading from, and + // wants to enqueue the maximum number of elements. We also need a spare element in each block to avoid the ambiguity + // between front == tail meaning "empty" and "full". + // So the effective number of slots that are guaranteed to be usable at any time is the block size - 1 times the + // number of blocks - 1. Solving for maxSize and applying a ceiling to the division gives us (after simplifying): + size_t initialBlockCount = (maxSize + MAX_BLOCK_SIZE * 2 - 3) / (MAX_BLOCK_SIZE - 1); + largestBlockSize = MAX_BLOCK_SIZE; + Block* lastBlock = nullptr; + for (size_t i = 0; i != initialBlockCount; ++i) { + auto block = make_block(largestBlockSize); + if (block == nullptr) { +#ifdef MOODYCAMEL_EXCEPTIONS_ENABLED + throw std::bad_alloc(); +#else + abort(); +#endif + } + if (firstBlock == nullptr) { + firstBlock = block; + } + else { + lastBlock->next = block; + } + lastBlock = block; + block->next = firstBlock; + } + } + else { + firstBlock = make_block(largestBlockSize); + if (firstBlock == nullptr) { +#ifdef MOODYCAMEL_EXCEPTIONS_ENABLED + throw std::bad_alloc(); +#else + abort(); +#endif + } + firstBlock->next = firstBlock; + } + frontBlock = firstBlock; + tailBlock = firstBlock; + + // Make sure the reader/writer threads will have the initialized memory setup above: + fence(memory_order_sync); + } + + // Note: The queue should not be accessed concurrently while it's + // being moved. It's up to the user to synchronize this. + AE_NO_TSAN ReaderWriterQueue(ReaderWriterQueue&& other) + : frontBlock(other.frontBlock.load()), + tailBlock(other.tailBlock.load()), + largestBlockSize(other.largestBlockSize) +#ifndef NDEBUG + ,enqueuing(false) + ,dequeuing(false) +#endif + { + other.largestBlockSize = 32; + Block* b = other.make_block(other.largestBlockSize); + if (b == nullptr) { +#ifdef MOODYCAMEL_EXCEPTIONS_ENABLED + throw std::bad_alloc(); +#else + abort(); +#endif + } + b->next = b; + other.frontBlock = b; + other.tailBlock = b; + } + + // Note: The queue should not be accessed concurrently while it's + // being moved. It's up to the user to synchronize this. + ReaderWriterQueue& operator=(ReaderWriterQueue&& other) AE_NO_TSAN + { + Block* b = frontBlock.load(); + frontBlock = other.frontBlock.load(); + other.frontBlock = b; + b = tailBlock.load(); + tailBlock = other.tailBlock.load(); + other.tailBlock = b; + std::swap(largestBlockSize, other.largestBlockSize); + return *this; + } + + // Note: The queue should not be accessed concurrently while it's + // being deleted. It's up to the user to synchronize this. + AE_NO_TSAN ~ReaderWriterQueue() + { + // Make sure we get the latest version of all variables from other CPUs: + fence(memory_order_sync); + + // Destroy any remaining objects in queue and free memory + Block* frontBlock_ = frontBlock; + Block* block = frontBlock_; + do { + Block* nextBlock = block->next; + size_t blockFront = block->front; + size_t blockTail = block->tail; + + for (size_t i = blockFront; i != blockTail; i = (i + 1) & block->sizeMask) { + auto element = reinterpret_cast(block->data + i * sizeof(T)); + element->~T(); + (void)element; + } + + auto rawBlock = block->rawThis; + block->~Block(); + std::free(rawBlock); + block = nextBlock; + } while (block != frontBlock_); + } + + + // Enqueues a copy of element if there is room in the queue. + // Returns true if the element was enqueued, false otherwise. + // Does not allocate memory. + AE_FORCEINLINE bool try_enqueue(T const& element) AE_NO_TSAN + { + return inner_enqueue(element); + } + + // Enqueues a moved copy of element if there is room in the queue. + // Returns true if the element was enqueued, false otherwise. + // Does not allocate memory. + AE_FORCEINLINE bool try_enqueue(T&& element) AE_NO_TSAN + { + return inner_enqueue(std::forward(element)); + } + +#if MOODYCAMEL_HAS_EMPLACE + // Like try_enqueue() but with emplace semantics (i.e. construct-in-place). + template + AE_FORCEINLINE bool try_emplace(Args&&... args) AE_NO_TSAN + { + return inner_enqueue(std::forward(args)...); + } +#endif + + // Enqueues a copy of element on the queue. + // Allocates an additional block of memory if needed. + // Only fails (returns false) if memory allocation fails. + AE_FORCEINLINE bool enqueue(T const& element) AE_NO_TSAN + { + return inner_enqueue(element); + } + + // Enqueues a moved copy of element on the queue. + // Allocates an additional block of memory if needed. + // Only fails (returns false) if memory allocation fails. + AE_FORCEINLINE bool enqueue(T&& element) AE_NO_TSAN + { + return inner_enqueue(std::forward(element)); + } + +#if MOODYCAMEL_HAS_EMPLACE + // Like enqueue() but with emplace semantics (i.e. construct-in-place). + template + AE_FORCEINLINE bool emplace(Args&&... args) AE_NO_TSAN + { + return inner_enqueue(std::forward(args)...); + } +#endif + + // Attempts to dequeue an element; if the queue is empty, + // returns false instead. If the queue has at least one element, + // moves front to result using operator=, then returns true. + template + bool try_dequeue(U& result) AE_NO_TSAN + { +#ifndef NDEBUG + ReentrantGuard guard(this->dequeuing); +#endif + + // High-level pseudocode: + // Remember where the tail block is + // If the front block has an element in it, dequeue it + // Else + // If front block was the tail block when we entered the function, return false + // Else advance to next block and dequeue the item there + + // Note that we have to use the value of the tail block from before we check if the front + // block is full or not, in case the front block is empty and then, before we check if the + // tail block is at the front block or not, the producer fills up the front block *and + // moves on*, which would make us skip a filled block. Seems unlikely, but was consistently + // reproducible in practice. + // In order to avoid overhead in the common case, though, we do a double-checked pattern + // where we have the fast path if the front block is not empty, then read the tail block, + // then re-read the front block and check if it's not empty again, then check if the tail + // block has advanced. + + Block* frontBlock_ = frontBlock.load(); + size_t blockTail = frontBlock_->localTail; + size_t blockFront = frontBlock_->front.load(); + + if (blockFront != blockTail || blockFront != (frontBlock_->localTail = frontBlock_->tail.load())) { + fence(memory_order_acquire); + + non_empty_front_block: + // Front block not empty, dequeue from here + auto element = reinterpret_cast(frontBlock_->data + blockFront * sizeof(T)); + result = std::move(*element); + element->~T(); + + blockFront = (blockFront + 1) & frontBlock_->sizeMask; + + fence(memory_order_release); + frontBlock_->front = blockFront; + } + else if (frontBlock_ != tailBlock.load()) { + fence(memory_order_acquire); + + frontBlock_ = frontBlock.load(); + blockTail = frontBlock_->localTail = frontBlock_->tail.load(); + blockFront = frontBlock_->front.load(); + fence(memory_order_acquire); + + if (blockFront != blockTail) { + // Oh look, the front block isn't empty after all + goto non_empty_front_block; + } + + // Front block is empty but there's another block ahead, advance to it + Block* nextBlock = frontBlock_->next; + // Don't need an acquire fence here since next can only ever be set on the tailBlock, + // and we're not the tailBlock, and we did an acquire earlier after reading tailBlock which + // ensures next is up-to-date on this CPU in case we recently were at tailBlock. + + size_t nextBlockFront = nextBlock->front.load(); + size_t nextBlockTail = nextBlock->localTail = nextBlock->tail.load(); + fence(memory_order_acquire); + + // Since the tailBlock is only ever advanced after being written to, + // we know there's for sure an element to dequeue on it + assert(nextBlockFront != nextBlockTail); + AE_UNUSED(nextBlockTail); + + // We're done with this block, let the producer use it if it needs + fence(memory_order_release); // Expose possibly pending changes to frontBlock->front from last dequeue + frontBlock = frontBlock_ = nextBlock; + + compiler_fence(memory_order_release); // Not strictly needed + + auto element = reinterpret_cast(frontBlock_->data + nextBlockFront * sizeof(T)); + + result = std::move(*element); + element->~T(); + + nextBlockFront = (nextBlockFront + 1) & frontBlock_->sizeMask; + + fence(memory_order_release); + frontBlock_->front = nextBlockFront; + } + else { + // No elements in current block and no other block to advance to + return false; + } + + return true; + } + + + // Returns a pointer to the front element in the queue (the one that + // would be removed next by a call to `try_dequeue` or `pop`). If the + // queue appears empty at the time the method is called, nullptr is + // returned instead. + // Must be called only from the consumer thread. + T* peek() AE_NO_TSAN + { +#ifndef NDEBUG + ReentrantGuard guard(this->dequeuing); +#endif + // See try_dequeue() for reasoning + + Block* frontBlock_ = frontBlock.load(); + size_t blockTail = frontBlock_->localTail; + size_t blockFront = frontBlock_->front.load(); + + if (blockFront != blockTail || blockFront != (frontBlock_->localTail = frontBlock_->tail.load())) { + fence(memory_order_acquire); + non_empty_front_block: + return reinterpret_cast(frontBlock_->data + blockFront * sizeof(T)); + } + else if (frontBlock_ != tailBlock.load()) { + fence(memory_order_acquire); + frontBlock_ = frontBlock.load(); + blockTail = frontBlock_->localTail = frontBlock_->tail.load(); + blockFront = frontBlock_->front.load(); + fence(memory_order_acquire); + + if (blockFront != blockTail) { + goto non_empty_front_block; + } + + Block* nextBlock = frontBlock_->next; + + size_t nextBlockFront = nextBlock->front.load(); + fence(memory_order_acquire); + + assert(nextBlockFront != nextBlock->tail.load()); + return reinterpret_cast(nextBlock->data + nextBlockFront * sizeof(T)); + } + + return nullptr; + } + + // Removes the front element from the queue, if any, without returning it. + // Returns true on success, or false if the queue appeared empty at the time + // `pop` was called. + bool pop() AE_NO_TSAN + { +#ifndef NDEBUG + ReentrantGuard guard(this->dequeuing); +#endif + // See try_dequeue() for reasoning + + Block* frontBlock_ = frontBlock.load(); + size_t blockTail = frontBlock_->localTail; + size_t blockFront = frontBlock_->front.load(); + + if (blockFront != blockTail || blockFront != (frontBlock_->localTail = frontBlock_->tail.load())) { + fence(memory_order_acquire); + + non_empty_front_block: + auto element = reinterpret_cast(frontBlock_->data + blockFront * sizeof(T)); + element->~T(); + + blockFront = (blockFront + 1) & frontBlock_->sizeMask; + + fence(memory_order_release); + frontBlock_->front = blockFront; + } + else if (frontBlock_ != tailBlock.load()) { + fence(memory_order_acquire); + frontBlock_ = frontBlock.load(); + blockTail = frontBlock_->localTail = frontBlock_->tail.load(); + blockFront = frontBlock_->front.load(); + fence(memory_order_acquire); + + if (blockFront != blockTail) { + goto non_empty_front_block; + } + + // Front block is empty but there's another block ahead, advance to it + Block* nextBlock = frontBlock_->next; + + size_t nextBlockFront = nextBlock->front.load(); + size_t nextBlockTail = nextBlock->localTail = nextBlock->tail.load(); + fence(memory_order_acquire); + + assert(nextBlockFront != nextBlockTail); + AE_UNUSED(nextBlockTail); + + fence(memory_order_release); + frontBlock = frontBlock_ = nextBlock; + + compiler_fence(memory_order_release); + + auto element = reinterpret_cast(frontBlock_->data + nextBlockFront * sizeof(T)); + element->~T(); + + nextBlockFront = (nextBlockFront + 1) & frontBlock_->sizeMask; + + fence(memory_order_release); + frontBlock_->front = nextBlockFront; + } + else { + // No elements in current block and no other block to advance to + return false; + } + + return true; + } + + // Returns the approximate number of items currently in the queue. + // Safe to call from both the producer and consumer threads. + inline size_t size_approx() const AE_NO_TSAN + { + size_t result = 0; + Block* frontBlock_ = frontBlock.load(); + Block* block = frontBlock_; + do { + fence(memory_order_acquire); + size_t blockFront = block->front.load(); + size_t blockTail = block->tail.load(); + result += (blockTail - blockFront) & block->sizeMask; + block = block->next.load(); + } while (block != frontBlock_); + return result; + } + + +private: + enum AllocationMode { CanAlloc, CannotAlloc }; + +#if MOODYCAMEL_HAS_EMPLACE + template + bool inner_enqueue(Args&&... args) AE_NO_TSAN +#else + template + bool inner_enqueue(U&& element) AE_NO_TSAN +#endif + { +#ifndef NDEBUG + ReentrantGuard guard(this->enqueuing); +#endif + + // High-level pseudocode (assuming we're allowed to alloc a new block): + // If room in tail block, add to tail + // Else check next block + // If next block is not the head block, enqueue on next block + // Else create a new block and enqueue there + // Advance tail to the block we just enqueued to + + Block* tailBlock_ = tailBlock.load(); + size_t blockFront = tailBlock_->localFront; + size_t blockTail = tailBlock_->tail.load(); + + size_t nextBlockTail = (blockTail + 1) & tailBlock_->sizeMask; + if (nextBlockTail != blockFront || nextBlockTail != (tailBlock_->localFront = tailBlock_->front.load())) { + fence(memory_order_acquire); + // This block has room for at least one more element + char* location = tailBlock_->data + blockTail * sizeof(T); +#if MOODYCAMEL_HAS_EMPLACE + new (location) T(std::forward(args)...); +#else + new (location) T(std::forward(element)); +#endif + + fence(memory_order_release); + tailBlock_->tail = nextBlockTail; + } + else { + fence(memory_order_acquire); + if (tailBlock_->next.load() != frontBlock) { + // Note that the reason we can't advance to the frontBlock and start adding new entries there + // is because if we did, then dequeue would stay in that block, eventually reading the new values, + // instead of advancing to the next full block (whose values were enqueued first and so should be + // consumed first). + + fence(memory_order_acquire); // Ensure we get latest writes if we got the latest frontBlock + + // tailBlock is full, but there's a free block ahead, use it + Block* tailBlockNext = tailBlock_->next.load(); + size_t nextBlockFront = tailBlockNext->localFront = tailBlockNext->front.load(); + nextBlockTail = tailBlockNext->tail.load(); + fence(memory_order_acquire); + + // This block must be empty since it's not the head block and we + // go through the blocks in a circle + assert(nextBlockFront == nextBlockTail); + tailBlockNext->localFront = nextBlockFront; + + char* location = tailBlockNext->data + nextBlockTail * sizeof(T); +#if MOODYCAMEL_HAS_EMPLACE + new (location) T(std::forward(args)...); +#else + new (location) T(std::forward(element)); +#endif + + tailBlockNext->tail = (nextBlockTail + 1) & tailBlockNext->sizeMask; + + fence(memory_order_release); + tailBlock = tailBlockNext; + } + else if (canAlloc == CanAlloc) { + // tailBlock is full and there's no free block ahead; create a new block + auto newBlockSize = largestBlockSize >= MAX_BLOCK_SIZE ? largestBlockSize : largestBlockSize * 2; + auto newBlock = make_block(newBlockSize); + if (newBlock == nullptr) { + // Could not allocate a block! + return false; + } + largestBlockSize = newBlockSize; + +#if MOODYCAMEL_HAS_EMPLACE + new (newBlock->data) T(std::forward(args)...); +#else + new (newBlock->data) T(std::forward(element)); +#endif + assert(newBlock->front == 0); + newBlock->tail = newBlock->localTail = 1; + + newBlock->next = tailBlock_->next.load(); + tailBlock_->next = newBlock; + + // Might be possible for the dequeue thread to see the new tailBlock->next + // *without* seeing the new tailBlock value, but this is OK since it can't + // advance to the next block until tailBlock is set anyway (because the only + // case where it could try to read the next is if it's already at the tailBlock, + // and it won't advance past tailBlock in any circumstance). + + fence(memory_order_release); + tailBlock = newBlock; + } + else if (canAlloc == CannotAlloc) { + // Would have had to allocate a new block to enqueue, but not allowed + return false; + } + else { + assert(false && "Should be unreachable code"); + return false; + } + } + + return true; + } + + + // Disable copying + ReaderWriterQueue(ReaderWriterQueue const&) { } + + // Disable assignment + ReaderWriterQueue& operator=(ReaderWriterQueue const&) { } + + + + AE_FORCEINLINE static size_t ceilToPow2(size_t x) + { + // From http://graphics.stanford.edu/~seander/bithacks.html#RoundUpPowerOf2 + --x; + x |= x >> 1; + x |= x >> 2; + x |= x >> 4; + for (size_t i = 1; i < sizeof(size_t); i <<= 1) { + x |= x >> (i << 3); + } + ++x; + return x; + } + + template + static AE_FORCEINLINE char* align_for(char* ptr) AE_NO_TSAN + { + const std::size_t alignment = std::alignment_of::value; + return ptr + (alignment - (reinterpret_cast(ptr) % alignment)) % alignment; + } +private: +#ifndef NDEBUG + struct ReentrantGuard + { + AE_NO_TSAN ReentrantGuard(bool& _inSection) + : inSection(_inSection) + { + assert(!inSection && "Concurrent (or re-entrant) enqueue or dequeue operation detected (only one thread at a time may hold the producer or consumer role)"); + inSection = true; + } + + AE_NO_TSAN ~ReentrantGuard() { inSection = false; } + + private: + ReentrantGuard& operator=(ReentrantGuard const&); + + private: + bool& inSection; + }; +#endif + + struct Block + { + // Avoid false-sharing by putting highly contended variables on their own cache lines + weak_atomic front; // (Atomic) Elements are read from here + size_t localTail; // An uncontended shadow copy of tail, owned by the consumer + + char cachelineFiller0[MOODYCAMEL_CACHE_LINE_SIZE - sizeof(weak_atomic) - sizeof(size_t)]; + weak_atomic tail; // (Atomic) Elements are enqueued here + size_t localFront; + + char cachelineFiller1[MOODYCAMEL_CACHE_LINE_SIZE - sizeof(weak_atomic) - sizeof(size_t)]; // next isn't very contended, but we don't want it on the same cache line as tail (which is) + weak_atomic next; // (Atomic) + + char* data; // Contents (on heap) are aligned to T's alignment + + const size_t sizeMask; + + + // size must be a power of two (and greater than 0) + AE_NO_TSAN Block(size_t const& _size, char* _rawThis, char* _data) + : front(0), localTail(0), tail(0), localFront(0), next(nullptr), data(_data), sizeMask(_size - 1), rawThis(_rawThis) + { + } + + private: + // C4512 - Assignment operator could not be generated + Block& operator=(Block const&); + + public: + char* rawThis; + }; + + + static Block* make_block(size_t capacity) AE_NO_TSAN + { + // Allocate enough memory for the block itself, as well as all the elements it will contain + auto size = sizeof(Block) + std::alignment_of::value - 1; + size += sizeof(T) * capacity + std::alignment_of::value - 1; + auto newBlockRaw = static_cast(std::malloc(size)); + if (newBlockRaw == nullptr) { + return nullptr; + } + + auto newBlockAligned = align_for(newBlockRaw); + auto newBlockData = align_for(newBlockAligned + sizeof(Block)); + return new (newBlockAligned) Block(capacity, newBlockRaw, newBlockData); + } + +private: + weak_atomic frontBlock; // (Atomic) Elements are enqueued to this block + + char cachelineFiller[MOODYCAMEL_CACHE_LINE_SIZE - sizeof(weak_atomic)]; + weak_atomic tailBlock; // (Atomic) Elements are dequeued from this block + + size_t largestBlockSize; + +#ifndef NDEBUG + bool enqueuing; + bool dequeuing; +#endif +}; + +// Like ReaderWriterQueue, but also providees blocking operations +template +class BlockingReaderWriterQueue +{ +private: + typedef ::moodycamel::ReaderWriterQueue ReaderWriterQueue; + +public: + explicit BlockingReaderWriterQueue(size_t maxSize = 15) AE_NO_TSAN + : inner(maxSize), sema(new spsc_sema::LightweightSemaphore()) + { } + + BlockingReaderWriterQueue(BlockingReaderWriterQueue&& other) AE_NO_TSAN + : inner(std::move(other.inner)), sema(std::move(other.sema)) + { } + + BlockingReaderWriterQueue& operator=(BlockingReaderWriterQueue&& other) AE_NO_TSAN + { + std::swap(sema, other.sema); + std::swap(inner, other.inner); + return *this; + } + + + // Enqueues a copy of element if there is room in the queue. + // Returns true if the element was enqueued, false otherwise. + // Does not allocate memory. + AE_FORCEINLINE bool try_enqueue(T const& element) AE_NO_TSAN + { + if (inner.try_enqueue(element)) { + sema->signal(); + return true; + } + return false; + } + + // Enqueues a moved copy of element if there is room in the queue. + // Returns true if the element was enqueued, false otherwise. + // Does not allocate memory. + AE_FORCEINLINE bool try_enqueue(T&& element) AE_NO_TSAN + { + if (inner.try_enqueue(std::forward(element))) { + sema->signal(); + return true; + } + return false; + } + + + // Enqueues a copy of element on the queue. + // Allocates an additional block of memory if needed. + // Only fails (returns false) if memory allocation fails. + AE_FORCEINLINE bool enqueue(T const& element) AE_NO_TSAN + { + if (inner.enqueue(element)) { + sema->signal(); + return true; + } + return false; + } + + // Enqueues a moved copy of element on the queue. + // Allocates an additional block of memory if needed. + // Only fails (returns false) if memory allocation fails. + AE_FORCEINLINE bool enqueue(T&& element) AE_NO_TSAN + { + if (inner.enqueue(std::forward(element))) { + sema->signal(); + return true; + } + return false; + } + + + // Attempts to dequeue an element; if the queue is empty, + // returns false instead. If the queue has at least one element, + // moves front to result using operator=, then returns true. + template + bool try_dequeue(U& result) AE_NO_TSAN + { + if (sema->tryWait()) { + bool success = inner.try_dequeue(result); + assert(success); + AE_UNUSED(success); + return true; + } + return false; + } + + + // Attempts to dequeue an element; if the queue is empty, + // waits until an element is available, then dequeues it. + template + void wait_dequeue(U& result) AE_NO_TSAN + { + sema->wait(); + bool success = inner.try_dequeue(result); + AE_UNUSED(result); + assert(success); + AE_UNUSED(success); + } + + + // Attempts to dequeue an element; if the queue is empty, + // waits until an element is available up to the specified timeout, + // then dequeues it and returns true, or returns false if the timeout + // expires before an element can be dequeued. + // Using a negative timeout indicates an indefinite timeout, + // and is thus functionally equivalent to calling wait_dequeue. + template + bool wait_dequeue_timed(U& result, std::int64_t timeout_usecs) AE_NO_TSAN + { + if (!sema->wait(timeout_usecs)) { + return false; + } + bool success = inner.try_dequeue(result); + AE_UNUSED(result); + assert(success); + AE_UNUSED(success); + return true; + } + + +#if __cplusplus > 199711L || _MSC_VER >= 1700 + // Attempts to dequeue an element; if the queue is empty, + // waits until an element is available up to the specified timeout, + // then dequeues it and returns true, or returns false if the timeout + // expires before an element can be dequeued. + // Using a negative timeout indicates an indefinite timeout, + // and is thus functionally equivalent to calling wait_dequeue. + template + inline bool wait_dequeue_timed(U& result, std::chrono::duration const& timeout) AE_NO_TSAN + { + return wait_dequeue_timed(result, std::chrono::duration_cast(timeout).count()); + } +#endif + + + // Returns a pointer to the front element in the queue (the one that + // would be removed next by a call to `try_dequeue` or `pop`). If the + // queue appears empty at the time the method is called, nullptr is + // returned instead. + // Must be called only from the consumer thread. + AE_FORCEINLINE T* peek() AE_NO_TSAN + { + return inner.peek(); + } + + // Removes the front element from the queue, if any, without returning it. + // Returns true on success, or false if the queue appeared empty at the time + // `pop` was called. + AE_FORCEINLINE bool pop() AE_NO_TSAN + { + if (sema->tryWait()) { + bool result = inner.pop(); + assert(result); + AE_UNUSED(result); + return true; + } + return false; + } + + // Returns the approximate number of items currently in the queue. + // Safe to call from both the producer and consumer threads. + AE_FORCEINLINE size_t size_approx() const AE_NO_TSAN + { + return sema->availableApprox(); + } + + +private: + // Disable copying & assignment + BlockingReaderWriterQueue(BlockingReaderWriterQueue const&) { } + BlockingReaderWriterQueue& operator=(BlockingReaderWriterQueue const&) { } + +private: + ReaderWriterQueue inner; + std::unique_ptr sema; +}; + +} // end namespace moodycamel + +#ifdef AE_VCPP +#pragma warning(pop) +#endif diff --git a/perf/remote_thr.cpp b/perf/remote_thr.cpp index 8378950cd9..6969e8c25f 100644 --- a/perf/remote_thr.cpp +++ b/perf/remote_thr.cpp @@ -31,12 +31,93 @@ #include #include #include +#include + +#include "readerwriterqueue.h" // keys are arbitrary but must match local_lat.cpp const char server_pubkey[] = "DX4nh=yUn{-9ugra0X3Src4SU-4xTgqxcYY.+raw_data); + + return true; + } + + static void + deallocate_msg (void *data_, + void *hint_) // producer thread: ZMQ background IO thread + { + ZmqMessagePool *pPool = reinterpret_cast (hint_); + + // recover the beginning of this msg_block: + uint8_t *data_ptr_ = (uint8_t *) data_; + msg_block_t *to_return = + (msg_block_t *) (data_ptr_ - SIZE_OF_CONTENT_T_USED_BY_ZMQ_VLM); + assert (to_return->canary == 0xAB); + + // produce a new free msg block: + pPool->m_free_list.enqueue (to_return); + } + + size_t size () const { return m_free_list.size_approx (); } + + private: + msg_block_t m_storage[MAX_ACTIVE_MESSAGES]; + moodycamel::ReaderWriterQueue m_free_list; +}; + + int main (int argc, char *argv[]) { const char *connect_to; @@ -104,6 +185,7 @@ int main (int argc, char *argv[]) return -1; } +#if 0 for (i = 0; i != message_count; i++) { rc = zmq_msg_init_size (&msg, message_size); if (rc != 0) { @@ -121,6 +203,32 @@ int main (int argc, char *argv[]) return -1; } } +#else + printf ("msg block size: %zu; max msg size: %d\n", sizeof (msg_block_t), + MAX_MESSAGE_SIZE); + ZmqMessagePool pool; + for (i = 0; i != message_count; i++) { + pool.allocate_msg (&msg, message_size); + + // to be fair when comparing the results generated by the other #if/#endif branch + // avoid any kind of initialization of message memory: + //memset (zmq_msg_data (&msg), message_size, 0xAB); + + rc = zmq_sendmsg (s, &msg, 0); + if (rc < 0) { + printf ("error in zmq_sendmsg: %s\n", zmq_strerror (errno)); + return -1; + } + rc = zmq_msg_close (&msg); + if (rc != 0) { + printf ("error in zmq_msg_close: %s\n", zmq_strerror (errno)); + return -1; + } + + //if ((i % 1000) == 0) + // printf ("mempool msg size: %zu\n", pool.size ()); + } +#endif rc = zmq_close (s); if (rc != 0) { diff --git a/src/msg.cpp b/src/msg.cpp index 5e32341bd0..dc1081c4c2 100644 --- a/src/msg.cpp +++ b/src/msg.cpp @@ -47,6 +47,9 @@ typedef char zmq_msg_size_check[2 * ((sizeof (zmq::msg_t) == sizeof (zmq_msg_t)) != 0) - 1]; +#define ALLOCATE_CONTENT_T_INSIDE_PROVIDED_BUFFER (1) + + bool zmq::msg_t::check () const { return _u.base.type >= type_min && _u.base.type <= type_max; @@ -166,15 +169,26 @@ int zmq::msg_t::init_data (void *data_, _u.lmsg.flags = 0; _u.lmsg.group[0] = '\0'; _u.lmsg.routing_id = 0; +#if ALLOCATE_CONTENT_T_INSIDE_PROVIDED_BUFFER + zmq_assert (size_ > sizeof (content_t)); + _u.lmsg.content = reinterpret_cast (data_); +#else _u.lmsg.content = static_cast (malloc (sizeof (content_t))); +#endif if (!_u.lmsg.content) { errno = ENOMEM; return -1; } +#if ALLOCATE_CONTENT_T_INSIDE_PROVIDED_BUFFER + uint8_t *data_bytes = (uint8_t *) data_; + _u.lmsg.content->data = data_bytes + sizeof (content_t); + _u.lmsg.content->size = size_ - sizeof (content_t); +#else _u.lmsg.content->data = data_; _u.lmsg.content->size = size_; +#endif _u.lmsg.content->ffn = ffn_; _u.lmsg.content->hint = hint_; new (&_u.lmsg.content->refcnt) zmq::atomic_counter_t (); @@ -228,11 +242,23 @@ int zmq::msg_t::close () // We used "placement new" operator to initialize the reference // counter so we call the destructor explicitly now. _u.lmsg.content->refcnt.~atomic_counter_t (); - +#if ALLOCATE_CONTENT_T_INSIDE_PROVIDED_BUFFER + // take a local copy since we are going to remove (through the user-provided deallocator) + // the whole malloc'ed buffer, including the content_t block itself! + // NOTE: this copy should not be strictly needed but it's here just to help debugging: + content_t content; + content.data = _u.lmsg.content->data; + content.size = _u.lmsg.content->size; + content.ffn = _u.lmsg.content->ffn; + content.hint = _u.lmsg.content->hint; + if (content.ffn) + content.ffn (content.data, content.hint); +#else if (_u.lmsg.content->ffn) _u.lmsg.content->ffn (_u.lmsg.content->data, _u.lmsg.content->hint); free (_u.lmsg.content); +#endif } } From a24f2af2579a7a402c9576e683f18d9b75ae7308 Mon Sep 17 00:00:00 2001 From: Francesco Montorsi Date: Tue, 13 Aug 2019 11:39:00 +0200 Subject: [PATCH 02/11] Allow to choose message sizes as well --- perf/generate_csv.sh | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/perf/generate_csv.sh b/perf/generate_csv.sh index d307f29e49..cf23ed5640 100755 --- a/perf/generate_csv.sh +++ b/perf/generate_csv.sh @@ -10,6 +10,7 @@ # export LOCAL_TEST_ENDPOINT="tcp://192.168.1.1:1234" # export REMOTE_TEST_ENDPOINT="tcp://192.168.1.2:1234" # export REMOTE_LIBZMQ_PATH="/home/fmontorsi/libzmq/perf" +# export MESSAGE_SIZE_LIST="8 16 32 64 128 210" # ./generate_csv.sh # @@ -22,7 +23,7 @@ LOCAL_TEST_ENDPOINT=${LOCAL_TEST_ENDPOINT:-tcp://192.168.1.1:1234} REMOTE_TEST_ENDPOINT=${REMOTE_TEST_ENDPOINT:-tcp://192.168.1.2:1234} # constant values: -MESSAGE_SIZE_LIST="8 16 32 64 128 256 512 1024 2048 4096 8192 16384 32768 65536 131072" +MESSAGE_SIZE_LIST="${MESSAGE_SIZE_LIST:-8 16 32 64 128 256 512 1024 2048 4096 8192 16384 32768 65536 131072}" OUTPUT_DIR="results" OUTPUT_FILE_PREFIX="results.txt" OUTPUT_FILE_CSV_PREFIX="results.csv" From 1bd2ae1530380937c69bd64262f42cb8710c6b7b Mon Sep 17 00:00:00 2001 From: Francesco Montorsi Date: Tue, 13 Aug 2019 11:50:49 +0200 Subject: [PATCH 03/11] Allow using env variables to do some basic overriding --- perf/generate_graphs.py | 49 ++++++++++++++++++++++++++++------------- 1 file changed, 34 insertions(+), 15 deletions(-) diff --git a/perf/generate_graphs.py b/perf/generate_graphs.py index 20651b7160..77fbb88ff3 100755 --- a/perf/generate_graphs.py +++ b/perf/generate_graphs.py @@ -1,19 +1,13 @@ #!/usr/bin/python3 # -# This script assumes that the set of CSV files produced by "generate_csv.sh" is provided as input -# and that locally there is the "results" folder. +# This script assumes that the set of CSV files produced by "generate_csv.sh" is provided as input. +# +# Usage example: +# export RESULT_DIRECTORY="./results" +# export TCP_LINK_SPEED_GBPS="10" # or 1 or 100 as you like +# ./generate_graphs.py # - -# results for TCP: -INPUT_FILE_PUSHPULL_TCP_THROUGHPUT="results/pushpull_tcp_thr_results.csv" -INPUT_FILE_REQREP_TCP_LATENCY="results/reqrep_tcp_lat_results.csv" -TCP_LINK_GPBS=100 - -# results for INPROC: -INPUT_FILE_PUSHPULL_INPROC_THROUGHPUT="results/pushpull_inproc_thr_results.csv" -INPUT_FILE_PUBSUBPROXY_INPROC_THROUGHPUT="results/pubsubproxy_inproc_thr_results.csv" - # dependencies # @@ -22,13 +16,15 @@ import matplotlib.pyplot as plt import numpy as np +import os # functions -def plot_throughput(csv_filename, title, is_tcp=False): +def plot_throughput(csv_filename, title, is_tcp=False, tcp_link_speed_gbps=10): message_size_bytes, message_count, pps, mbps = np.loadtxt(csv_filename, delimiter=',', unpack=True) + print("Generating PNG image file [%s] from CSV results '%s'" % (title, csv_filename)) fig, ax1 = plt.subplots() # PPS axis @@ -44,7 +40,7 @@ def plot_throughput(csv_filename, title, is_tcp=False): ax2.set_ylabel('Throughput [Gb/s]', color=color) ax2.semilogx(message_size_bytes, mbps / 1e3, label='Throughput [Gb/s]', marker='o') if is_tcp: - ax2.set_yticks(np.arange(0, TCP_LINK_GPBS + 1, TCP_LINK_GPBS/10)) + ax2.set_yticks(np.arange(0, tcp_link_speed_gbps + 1, tcp_link_speed_gbps/10)) ax2.tick_params(axis='y', labelcolor=color) ax2.grid(True) @@ -55,6 +51,8 @@ def plot_throughput(csv_filename, title, is_tcp=False): def plot_latency(csv_filename, title): message_size_bytes, message_count, lat = np.loadtxt(csv_filename, delimiter=',', unpack=True) + + print("Generating PNG image file [%s] from CSV results '%s'" % (title, csv_filename)) plt.semilogx(message_size_bytes, lat, label='Latency [us]', marker='o') plt.xlabel('Message size [B]') @@ -67,7 +65,28 @@ def plot_latency(csv_filename, title): # main -plot_throughput(INPUT_FILE_PUSHPULL_TCP_THROUGHPUT, 'ZeroMQ PUSH/PULL socket throughput, TCP transport', is_tcp=True) +try: + result_dir = os.environ['RESULT_DIRECTORY'] +except: + result_dir = "results" # default value + +try: + tcp_link_speed_gbps = os.environ['TCP_LINK_SPEED_GBPS'] +except: + tcp_link_speed_gbps = "10" # default value + + + +# result files for TCP: +INPUT_FILE_PUSHPULL_TCP_THROUGHPUT = result_dir + "/pushpull_tcp_thr_results.csv" +INPUT_FILE_REQREP_TCP_LATENCY = result_dir + "/reqrep_tcp_lat_results.csv" + +# results for INPROC: +INPUT_FILE_PUSHPULL_INPROC_THROUGHPUT = result_dir + "/pushpull_inproc_thr_results.csv" +INPUT_FILE_PUBSUBPROXY_INPROC_THROUGHPUT = result_dir + "/pubsubproxy_inproc_thr_results.csv" + +# generate plots +plot_throughput(INPUT_FILE_PUSHPULL_TCP_THROUGHPUT, 'ZeroMQ PUSH/PULL socket throughput, TCP transport', is_tcp=True, tcp_link_speed_gbps=tcp_link_speed_gbps) plot_throughput(INPUT_FILE_PUSHPULL_INPROC_THROUGHPUT, 'ZeroMQ PUSH/PULL socket throughput, INPROC transport') plot_throughput(INPUT_FILE_PUBSUBPROXY_INPROC_THROUGHPUT, 'ZeroMQ PUB/SUB PROXY socket throughput, INPROC transport') plot_latency(INPUT_FILE_REQREP_TCP_LATENCY, 'ZeroMQ REQ/REP socket latency, TCP transport') From 252e8d449c6c40919f81de351d34be8a02af6ed2 Mon Sep 17 00:00:00 2001 From: Francesco Montorsi Date: Tue, 13 Aug 2019 11:51:39 +0200 Subject: [PATCH 04/11] fix typo --- perf/generate_graphs.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/perf/generate_graphs.py b/perf/generate_graphs.py index 77fbb88ff3..c323e4cce6 100755 --- a/perf/generate_graphs.py +++ b/perf/generate_graphs.py @@ -71,9 +71,9 @@ def plot_latency(csv_filename, title): result_dir = "results" # default value try: - tcp_link_speed_gbps = os.environ['TCP_LINK_SPEED_GBPS'] + tcp_link_speed_gbps = int(os.environ['TCP_LINK_SPEED_GBPS']) except: - tcp_link_speed_gbps = "10" # default value + tcp_link_speed_gbps = 10 # default value From 4a3079560b6be68a9a1ad5291ed469cf1b35379a Mon Sep 17 00:00:00 2001 From: Francesco Montorsi Date: Tue, 13 Aug 2019 12:05:50 +0200 Subject: [PATCH 05/11] add TCP kernel socket buffer setting --- perf/generate_csv.sh | 30 ++++++++++++++++++++++++++++++ 1 file changed, 30 insertions(+) diff --git a/perf/generate_csv.sh b/perf/generate_csv.sh index cf23ed5640..da8ff0a4cd 100755 --- a/perf/generate_csv.sh +++ b/perf/generate_csv.sh @@ -48,6 +48,35 @@ function verify_ssh() echo "SSH connection to the remote $REMOTE_IP_SSH is working fine." } +function set_reproducible_tcp_kernel_buff_size() +{ + sysctl -w net.core.rmem_max=8388608 && \ + sysctl -w net.core.wmem_max=8388608 && \ + sysctl -w net.core.rmem_default=65536 && \ + sysctl -w net.core.wmem_default=65536 && \ + sysctl -w net.ipv4.tcp_rmem='4096 87380 8388608' && \ + sysctl -w net.ipv4.tcp_wmem='4096 65536 8388608' && \ + sysctl -w net.ipv4.tcp_mem='8388608 8388608 8388608' && \ + sysctl -w net.ipv4.route.flush=1 + if [ $? -ne 0 ]; then + echo "Failed setting kernel socket buffer sizes LOCALLY" + exit 2 + fi + + ssh $REMOTE_IP_SSH "sysctl -w net.core.rmem_max=8388608 && \ + sysctl -w net.core.wmem_max=8388608 && \ + sysctl -w net.core.rmem_default=65536 && \ + sysctl -w net.core.wmem_default=65536 && \ + sysctl -w net.ipv4.tcp_rmem='4096 87380 8388608' && \ + sysctl -w net.ipv4.tcp_wmem='4096 65536 8388608' && \ + sysctl -w net.ipv4.tcp_mem='8388608 8388608 8388608' && \ + sysctl -w net.ipv4.route.flush=1" + if [ $? -ne 0 ]; then + echo "Failed setting kernel socket buffer sizes on the REMOTE system $REMOTE_IP_SSH" + exit 2 + fi +} + function run_remote_perf_util() { local MESSAGE_SIZE_BYTES="$1" @@ -112,6 +141,7 @@ function generate_output_file() # main: verify_ssh +set_reproducible_tcp_kernel_buff_size THROUGHPUT_CSV_HEADER_LINE="# message_size,message_count,PPS[msg/s],throughput[Mb/s]" From 00e514e2c9b8ae0373a2b9d0e594310efd152fe5 Mon Sep 17 00:00:00 2001 From: Francesco Montorsi Date: Thu, 29 Aug 2019 00:39:31 +0200 Subject: [PATCH 06/11] First implementation of global memory pool for ZMQ --- Makefile.am | 2 + include/atomicops.h | 676 ------- include/readerwriterqueue.h | 906 --------- include/zmq.h | 14 + perf/remote_thr.cpp | 124 +- src/allocator.cpp | 97 + src/allocator.hpp | 181 ++ src/concurrentqueue.h | 3636 +++++++++++++++++++++++++++++++++++ src/ctx.cpp | 16 + src/ctx.hpp | 4 + src/msg.cpp | 79 +- src/msg.hpp | 6 +- src/zmq.cpp | 38 + 13 files changed, 4058 insertions(+), 1721 deletions(-) delete mode 100644 include/atomicops.h delete mode 100644 include/readerwriterqueue.h create mode 100644 src/allocator.cpp create mode 100644 src/allocator.hpp create mode 100644 src/concurrentqueue.h diff --git a/Makefile.am b/Makefile.am index 4c4abc4415..e81a4ca7a9 100644 --- a/Makefile.am +++ b/Makefile.am @@ -22,6 +22,8 @@ include_HEADERS = \ src_libzmq_la_SOURCES = \ src/address.cpp \ src/address.hpp \ + src/allocator.cpp \ + src/allocator.hpp \ src/array.hpp \ src/atomic_counter.hpp \ src/atomic_ptr.hpp \ diff --git a/include/atomicops.h b/include/atomicops.h deleted file mode 100644 index 4fd1748293..0000000000 --- a/include/atomicops.h +++ /dev/null @@ -1,676 +0,0 @@ -// ©2013-2016 Cameron Desrochers. -// Distributed under the simplified BSD license (see the license file that -// should have come with this header). -// Uses Jeff Preshing's semaphore implementation (under the terms of its -// separate zlib license, embedded below). - -#pragma once - -// Provides portable (VC++2010+, Intel ICC 13, GCC 4.7+, and anything C++11 compliant) implementation -// of low-level memory barriers, plus a few semi-portable utility macros (for inlining and alignment). -// Also has a basic atomic type (limited to hardware-supported atomics with no memory ordering guarantees). -// Uses the AE_* prefix for macros (historical reasons), and the "moodycamel" namespace for symbols. - -#include -#include -#include -#include -#include - -// Platform detection -#if defined(__INTEL_COMPILER) -#define AE_ICC -#elif defined(_MSC_VER) -#define AE_VCPP -#elif defined(__GNUC__) -#define AE_GCC -#endif - -#if defined(_M_IA64) || defined(__ia64__) -#define AE_ARCH_IA64 -#elif defined(_WIN64) || defined(__amd64__) || defined(_M_X64) || defined(__x86_64__) -#define AE_ARCH_X64 -#elif defined(_M_IX86) || defined(__i386__) -#define AE_ARCH_X86 -#elif defined(_M_PPC) || defined(__powerpc__) -#define AE_ARCH_PPC -#else -#define AE_ARCH_UNKNOWN -#endif - - -// AE_UNUSED -#define AE_UNUSED(x) ((void)x) - -// AE_NO_TSAN -#if defined(__has_feature) -#if __has_feature(thread_sanitizer) -#define AE_NO_TSAN __attribute__((no_sanitize("thread"))) -#else -#define AE_NO_TSAN -#endif -#else -#define AE_NO_TSAN -#endif - - -// AE_FORCEINLINE -#if defined(AE_VCPP) || defined(AE_ICC) -#define AE_FORCEINLINE __forceinline -#elif defined(AE_GCC) -//#define AE_FORCEINLINE __attribute__((always_inline)) -#define AE_FORCEINLINE inline -#else -#define AE_FORCEINLINE inline -#endif - - -// AE_ALIGN -#if defined(AE_VCPP) || defined(AE_ICC) -#define AE_ALIGN(x) __declspec(align(x)) -#elif defined(AE_GCC) -#define AE_ALIGN(x) __attribute__((aligned(x))) -#else -// Assume GCC compliant syntax... -#define AE_ALIGN(x) __attribute__((aligned(x))) -#endif - - -// Portable atomic fences implemented below: - -namespace moodycamel { - -enum memory_order { - memory_order_relaxed, - memory_order_acquire, - memory_order_release, - memory_order_acq_rel, - memory_order_seq_cst, - - // memory_order_sync: Forces a full sync: - // #LoadLoad, #LoadStore, #StoreStore, and most significantly, #StoreLoad - memory_order_sync = memory_order_seq_cst -}; - -} // end namespace moodycamel - -#if (defined(AE_VCPP) && (_MSC_VER < 1700 || defined(__cplusplus_cli))) || (defined(AE_ICC) && __INTEL_COMPILER < 1600) -// VS2010 and ICC13 don't support std::atomic_*_fence, implement our own fences - -#include - -#if defined(AE_ARCH_X64) || defined(AE_ARCH_X86) -#define AeFullSync _mm_mfence -#define AeLiteSync _mm_mfence -#elif defined(AE_ARCH_IA64) -#define AeFullSync __mf -#define AeLiteSync __mf -#elif defined(AE_ARCH_PPC) -#include -#define AeFullSync __sync -#define AeLiteSync __lwsync -#endif - - -#ifdef AE_VCPP -#pragma warning(push) -#pragma warning(disable: 4365) // Disable erroneous 'conversion from long to unsigned int, signed/unsigned mismatch' error when using `assert` -#ifdef __cplusplus_cli -#pragma managed(push, off) -#endif -#endif - -namespace moodycamel { - -AE_FORCEINLINE void compiler_fence(memory_order order) AE_NO_TSAN -{ - switch (order) { - case memory_order_relaxed: break; - case memory_order_acquire: _ReadBarrier(); break; - case memory_order_release: _WriteBarrier(); break; - case memory_order_acq_rel: _ReadWriteBarrier(); break; - case memory_order_seq_cst: _ReadWriteBarrier(); break; - default: assert(false); - } -} - -// x86/x64 have a strong memory model -- all loads and stores have -// acquire and release semantics automatically (so only need compiler -// barriers for those). -#if defined(AE_ARCH_X86) || defined(AE_ARCH_X64) -AE_FORCEINLINE void fence(memory_order order) AE_NO_TSAN -{ - switch (order) { - case memory_order_relaxed: break; - case memory_order_acquire: _ReadBarrier(); break; - case memory_order_release: _WriteBarrier(); break; - case memory_order_acq_rel: _ReadWriteBarrier(); break; - case memory_order_seq_cst: - _ReadWriteBarrier(); - AeFullSync(); - _ReadWriteBarrier(); - break; - default: assert(false); - } -} -#else -AE_FORCEINLINE void fence(memory_order order) AE_NO_TSAN -{ - // Non-specialized arch, use heavier memory barriers everywhere just in case :-( - switch (order) { - case memory_order_relaxed: - break; - case memory_order_acquire: - _ReadBarrier(); - AeLiteSync(); - _ReadBarrier(); - break; - case memory_order_release: - _WriteBarrier(); - AeLiteSync(); - _WriteBarrier(); - break; - case memory_order_acq_rel: - _ReadWriteBarrier(); - AeLiteSync(); - _ReadWriteBarrier(); - break; - case memory_order_seq_cst: - _ReadWriteBarrier(); - AeFullSync(); - _ReadWriteBarrier(); - break; - default: assert(false); - } -} -#endif -} // end namespace moodycamel -#else -// Use standard library of atomics -#include - -namespace moodycamel { - -AE_FORCEINLINE void compiler_fence(memory_order order) AE_NO_TSAN -{ - switch (order) { - case memory_order_relaxed: break; - case memory_order_acquire: std::atomic_signal_fence(std::memory_order_acquire); break; - case memory_order_release: std::atomic_signal_fence(std::memory_order_release); break; - case memory_order_acq_rel: std::atomic_signal_fence(std::memory_order_acq_rel); break; - case memory_order_seq_cst: std::atomic_signal_fence(std::memory_order_seq_cst); break; - default: assert(false); - } -} - -AE_FORCEINLINE void fence(memory_order order) AE_NO_TSAN -{ - switch (order) { - case memory_order_relaxed: break; - case memory_order_acquire: std::atomic_thread_fence(std::memory_order_acquire); break; - case memory_order_release: std::atomic_thread_fence(std::memory_order_release); break; - case memory_order_acq_rel: std::atomic_thread_fence(std::memory_order_acq_rel); break; - case memory_order_seq_cst: std::atomic_thread_fence(std::memory_order_seq_cst); break; - default: assert(false); - } -} - -} // end namespace moodycamel - -#endif - - -#if !defined(AE_VCPP) || (_MSC_VER >= 1700 && !defined(__cplusplus_cli)) -#define AE_USE_STD_ATOMIC_FOR_WEAK_ATOMIC -#endif - -#ifdef AE_USE_STD_ATOMIC_FOR_WEAK_ATOMIC -#include -#endif -#include - -// WARNING: *NOT* A REPLACEMENT FOR std::atomic. READ CAREFULLY: -// Provides basic support for atomic variables -- no memory ordering guarantees are provided. -// The guarantee of atomicity is only made for types that already have atomic load and store guarantees -// at the hardware level -- on most platforms this generally means aligned pointers and integers (only). -namespace moodycamel { -template -class weak_atomic -{ -public: - AE_NO_TSAN weak_atomic() { } -#ifdef AE_VCPP -#pragma warning(push) -#pragma warning(disable: 4100) // Get rid of (erroneous) 'unreferenced formal parameter' warning -#endif - template AE_NO_TSAN weak_atomic(U&& x) : value(std::forward(x)) { } -#ifdef __cplusplus_cli - // Work around bug with universal reference/nullptr combination that only appears when /clr is on - AE_NO_TSAN weak_atomic(nullptr_t) : value(nullptr) { } -#endif - AE_NO_TSAN weak_atomic(weak_atomic const& other) : value(other.load()) { } - AE_NO_TSAN weak_atomic(weak_atomic&& other) : value(std::move(other.load())) { } -#ifdef AE_VCPP -#pragma warning(pop) -#endif - - AE_FORCEINLINE operator T() const AE_NO_TSAN { return load(); } - - -#ifndef AE_USE_STD_ATOMIC_FOR_WEAK_ATOMIC - template AE_FORCEINLINE weak_atomic const& operator=(U&& x) AE_NO_TSAN { value = std::forward(x); return *this; } - AE_FORCEINLINE weak_atomic const& operator=(weak_atomic const& other) AE_NO_TSAN { value = other.value; return *this; } - - AE_FORCEINLINE T load() const AE_NO_TSAN { return value; } - - AE_FORCEINLINE T fetch_add_acquire(T increment) AE_NO_TSAN - { -#if defined(AE_ARCH_X64) || defined(AE_ARCH_X86) - if (sizeof(T) == 4) return _InterlockedExchangeAdd((long volatile*)&value, (long)increment); -#if defined(_M_AMD64) - else if (sizeof(T) == 8) return _InterlockedExchangeAdd64((long long volatile*)&value, (long long)increment); -#endif -#else -#error Unsupported platform -#endif - assert(false && "T must be either a 32 or 64 bit type"); - return value; - } - - AE_FORCEINLINE T fetch_add_release(T increment) AE_NO_TSAN - { -#if defined(AE_ARCH_X64) || defined(AE_ARCH_X86) - if (sizeof(T) == 4) return _InterlockedExchangeAdd((long volatile*)&value, (long)increment); -#if defined(_M_AMD64) - else if (sizeof(T) == 8) return _InterlockedExchangeAdd64((long long volatile*)&value, (long long)increment); -#endif -#else -#error Unsupported platform -#endif - assert(false && "T must be either a 32 or 64 bit type"); - return value; - } -#else - template - AE_FORCEINLINE weak_atomic const& operator=(U&& x) AE_NO_TSAN - { - value.store(std::forward(x), std::memory_order_relaxed); - return *this; - } - - AE_FORCEINLINE weak_atomic const& operator=(weak_atomic const& other) AE_NO_TSAN - { - value.store(other.value.load(std::memory_order_relaxed), std::memory_order_relaxed); - return *this; - } - - AE_FORCEINLINE T load() const AE_NO_TSAN { return value.load(std::memory_order_relaxed); } - - AE_FORCEINLINE T fetch_add_acquire(T increment) AE_NO_TSAN - { - return value.fetch_add(increment, std::memory_order_acquire); - } - - AE_FORCEINLINE T fetch_add_release(T increment) AE_NO_TSAN - { - return value.fetch_add(increment, std::memory_order_release); - } -#endif - - -private: -#ifndef AE_USE_STD_ATOMIC_FOR_WEAK_ATOMIC - // No std::atomic support, but still need to circumvent compiler optimizations. - // `volatile` will make memory access slow, but is guaranteed to be reliable. - volatile T value; -#else - std::atomic value; -#endif -}; - -} // end namespace moodycamel - - - -// Portable single-producer, single-consumer semaphore below: - -#if defined(_WIN32) -// Avoid including windows.h in a header; we only need a handful of -// items, so we'll redeclare them here (this is relatively safe since -// the API generally has to remain stable between Windows versions). -// I know this is an ugly hack but it still beats polluting the global -// namespace with thousands of generic names or adding a .cpp for nothing. -extern "C" { - struct _SECURITY_ATTRIBUTES; - __declspec(dllimport) void* __stdcall CreateSemaphoreW(_SECURITY_ATTRIBUTES* lpSemaphoreAttributes, long lInitialCount, long lMaximumCount, const wchar_t* lpName); - __declspec(dllimport) int __stdcall CloseHandle(void* hObject); - __declspec(dllimport) unsigned long __stdcall WaitForSingleObject(void* hHandle, unsigned long dwMilliseconds); - __declspec(dllimport) int __stdcall ReleaseSemaphore(void* hSemaphore, long lReleaseCount, long* lpPreviousCount); -} -#elif defined(__MACH__) -#include -#elif defined(__unix__) -#include -#endif - -namespace moodycamel -{ - // Code in the spsc_sema namespace below is an adaptation of Jeff Preshing's - // portable + lightweight semaphore implementations, originally from - // https://github.com/preshing/cpp11-on-multicore/blob/master/common/sema.h - // LICENSE: - // Copyright (c) 2015 Jeff Preshing - // - // This software is provided 'as-is', without any express or implied - // warranty. In no event will the authors be held liable for any damages - // arising from the use of this software. - // - // Permission is granted to anyone to use this software for any purpose, - // including commercial applications, and to alter it and redistribute it - // freely, subject to the following restrictions: - // - // 1. The origin of this software must not be misrepresented; you must not - // claim that you wrote the original software. If you use this software - // in a product, an acknowledgement in the product documentation would be - // appreciated but is not required. - // 2. Altered source versions must be plainly marked as such, and must not be - // misrepresented as being the original software. - // 3. This notice may not be removed or altered from any source distribution. - namespace spsc_sema - { -#if defined(_WIN32) - class Semaphore - { - private: - void* m_hSema; - - Semaphore(const Semaphore& other); - Semaphore& operator=(const Semaphore& other); - - public: - AE_NO_TSAN Semaphore(int initialCount = 0) - { - assert(initialCount >= 0); - const long maxLong = 0x7fffffff; - m_hSema = CreateSemaphoreW(nullptr, initialCount, maxLong, nullptr); - } - - AE_NO_TSAN ~Semaphore() - { - CloseHandle(m_hSema); - } - - void wait() AE_NO_TSAN - { - const unsigned long infinite = 0xffffffff; - WaitForSingleObject(m_hSema, infinite); - } - - bool try_wait() AE_NO_TSAN - { - const unsigned long RC_WAIT_TIMEOUT = 0x00000102; - return WaitForSingleObject(m_hSema, 0) != RC_WAIT_TIMEOUT; - } - - bool timed_wait(std::uint64_t usecs) AE_NO_TSAN - { - const unsigned long RC_WAIT_TIMEOUT = 0x00000102; - return WaitForSingleObject(m_hSema, (unsigned long)(usecs / 1000)) != RC_WAIT_TIMEOUT; - } - - void signal(int count = 1) AE_NO_TSAN - { - ReleaseSemaphore(m_hSema, count, nullptr); - } - }; -#elif defined(__MACH__) - //--------------------------------------------------------- - // Semaphore (Apple iOS and OSX) - // Can't use POSIX semaphores due to http://lists.apple.com/archives/darwin-kernel/2009/Apr/msg00010.html - //--------------------------------------------------------- - class Semaphore - { - private: - semaphore_t m_sema; - - Semaphore(const Semaphore& other); - Semaphore& operator=(const Semaphore& other); - - public: - AE_NO_TSAN Semaphore(int initialCount = 0) - { - assert(initialCount >= 0); - semaphore_create(mach_task_self(), &m_sema, SYNC_POLICY_FIFO, initialCount); - } - - AE_NO_TSAN ~Semaphore() - { - semaphore_destroy(mach_task_self(), m_sema); - } - - void wait() AE_NO_TSAN - { - semaphore_wait(m_sema); - } - - bool try_wait() AE_NO_TSAN - { - return timed_wait(0); - } - - bool timed_wait(std::int64_t timeout_usecs) AE_NO_TSAN - { - mach_timespec_t ts; - ts.tv_sec = static_cast(timeout_usecs / 1000000); - ts.tv_nsec = (timeout_usecs % 1000000) * 1000; - - // added in OSX 10.10: https://developer.apple.com/library/prerelease/mac/documentation/General/Reference/APIDiffsMacOSX10_10SeedDiff/modules/Darwin.html - kern_return_t rc = semaphore_timedwait(m_sema, ts); - - return rc != KERN_OPERATION_TIMED_OUT && rc != KERN_ABORTED; - } - - void signal() AE_NO_TSAN - { - semaphore_signal(m_sema); - } - - void signal(int count) AE_NO_TSAN - { - while (count-- > 0) - { - semaphore_signal(m_sema); - } - } - }; -#elif defined(__unix__) - //--------------------------------------------------------- - // Semaphore (POSIX, Linux) - //--------------------------------------------------------- - class Semaphore - { - private: - sem_t m_sema; - - Semaphore(const Semaphore& other); - Semaphore& operator=(const Semaphore& other); - - public: - AE_NO_TSAN Semaphore(int initialCount = 0) - { - assert(initialCount >= 0); - sem_init(&m_sema, 0, initialCount); - } - - AE_NO_TSAN ~Semaphore() - { - sem_destroy(&m_sema); - } - - void wait() AE_NO_TSAN - { - // http://stackoverflow.com/questions/2013181/gdb-causes-sem-wait-to-fail-with-eintr-error - int rc; - do - { - rc = sem_wait(&m_sema); - } - while (rc == -1 && errno == EINTR); - } - - bool try_wait() AE_NO_TSAN - { - int rc; - do { - rc = sem_trywait(&m_sema); - } while (rc == -1 && errno == EINTR); - return !(rc == -1 && errno == EAGAIN); - } - - bool timed_wait(std::uint64_t usecs) AE_NO_TSAN - { - struct timespec ts; - const int usecs_in_1_sec = 1000000; - const int nsecs_in_1_sec = 1000000000; - clock_gettime(CLOCK_REALTIME, &ts); - ts.tv_sec += usecs / usecs_in_1_sec; - ts.tv_nsec += (usecs % usecs_in_1_sec) * 1000; - // sem_timedwait bombs if you have more than 1e9 in tv_nsec - // so we have to clean things up before passing it in - if (ts.tv_nsec >= nsecs_in_1_sec) { - ts.tv_nsec -= nsecs_in_1_sec; - ++ts.tv_sec; - } - - int rc; - do { - rc = sem_timedwait(&m_sema, &ts); - } while (rc == -1 && errno == EINTR); - return !(rc == -1 && errno == ETIMEDOUT); - } - - void signal() AE_NO_TSAN - { - sem_post(&m_sema); - } - - void signal(int count) AE_NO_TSAN - { - while (count-- > 0) - { - sem_post(&m_sema); - } - } - }; -#else -#error Unsupported platform! (No semaphore wrapper available) -#endif - - //--------------------------------------------------------- - // LightweightSemaphore - //--------------------------------------------------------- - class LightweightSemaphore - { - public: - typedef std::make_signed::type ssize_t; - - private: - weak_atomic m_count; - Semaphore m_sema; - - bool waitWithPartialSpinning(std::int64_t timeout_usecs = -1) AE_NO_TSAN - { - ssize_t oldCount; - // Is there a better way to set the initial spin count? - // If we lower it to 1000, testBenaphore becomes 15x slower on my Core i7-5930K Windows PC, - // as threads start hitting the kernel semaphore. - int spin = 10000; - while (--spin >= 0) - { - if (m_count.load() > 0) - { - m_count.fetch_add_acquire(-1); - return true; - } - compiler_fence(memory_order_acquire); // Prevent the compiler from collapsing the loop. - } - oldCount = m_count.fetch_add_acquire(-1); - if (oldCount > 0) - return true; - if (timeout_usecs < 0) - { - m_sema.wait(); - return true; - } - if (m_sema.timed_wait(timeout_usecs)) - return true; - // At this point, we've timed out waiting for the semaphore, but the - // count is still decremented indicating we may still be waiting on - // it. So we have to re-adjust the count, but only if the semaphore - // wasn't signaled enough times for us too since then. If it was, we - // need to release the semaphore too. - while (true) - { - oldCount = m_count.fetch_add_release(1); - if (oldCount < 0) - return false; // successfully restored things to the way they were - // Oh, the producer thread just signaled the semaphore after all. Try again: - oldCount = m_count.fetch_add_acquire(-1); - if (oldCount > 0 && m_sema.try_wait()) - return true; - } - } - - public: - AE_NO_TSAN LightweightSemaphore(ssize_t initialCount = 0) : m_count(initialCount) - { - assert(initialCount >= 0); - } - - bool tryWait() AE_NO_TSAN - { - if (m_count.load() > 0) - { - m_count.fetch_add_acquire(-1); - return true; - } - return false; - } - - void wait() AE_NO_TSAN - { - if (!tryWait()) - waitWithPartialSpinning(); - } - - bool wait(std::int64_t timeout_usecs) AE_NO_TSAN - { - return tryWait() || waitWithPartialSpinning(timeout_usecs); - } - - void signal(ssize_t count = 1) AE_NO_TSAN - { - assert(count >= 0); - ssize_t oldCount = m_count.fetch_add_release(count); - assert(oldCount >= -1); - if (oldCount < 0) - { - m_sema.signal(1); - } - } - - ssize_t availableApprox() const AE_NO_TSAN - { - ssize_t count = m_count.load(); - return count > 0 ? count : 0; - } - }; - } // end namespace spsc_sema -} // end namespace moodycamel - -#if defined(AE_VCPP) && (_MSC_VER < 1700 || defined(__cplusplus_cli)) -#pragma warning(pop) -#ifdef __cplusplus_cli -#pragma managed(pop) -#endif -#endif diff --git a/include/readerwriterqueue.h b/include/readerwriterqueue.h deleted file mode 100644 index 071147c3e1..0000000000 --- a/include/readerwriterqueue.h +++ /dev/null @@ -1,906 +0,0 @@ -// ©2013-2016 Cameron Desrochers. -// Distributed under the simplified BSD license (see the license file that -// should have come with this header). - -#pragma once - -#include "atomicops.h" -#include -#include -#include -#include -#include -#include -#include // For malloc/free/abort & size_t -#include -#if __cplusplus > 199711L || _MSC_VER >= 1700 // C++11 or VS2012 -#include -#endif - - -// A lock-free queue for a single-consumer, single-producer architecture. -// The queue is also wait-free in the common path (except if more memory -// needs to be allocated, in which case malloc is called). -// Allocates memory sparingly (O(lg(n) times, amortized), and only once if -// the original maximum size estimate is never exceeded. -// Tested on x86/x64 processors, but semantics should be correct for all -// architectures (given the right implementations in atomicops.h), provided -// that aligned integer and pointer accesses are naturally atomic. -// Note that there should only be one consumer thread and producer thread; -// Switching roles of the threads, or using multiple consecutive threads for -// one role, is not safe unless properly synchronized. -// Using the queue exclusively from one thread is fine, though a bit silly. - -#ifndef MOODYCAMEL_CACHE_LINE_SIZE -#define MOODYCAMEL_CACHE_LINE_SIZE 64 -#endif - -#ifndef MOODYCAMEL_EXCEPTIONS_ENABLED -#if (defined(_MSC_VER) && defined(_CPPUNWIND)) || (defined(__GNUC__) && defined(__EXCEPTIONS)) || (!defined(_MSC_VER) && !defined(__GNUC__)) -#define MOODYCAMEL_EXCEPTIONS_ENABLED -#endif -#endif - -#ifndef MOODYCAMEL_HAS_EMPLACE -#if !defined(_MSC_VER) || _MSC_VER >= 1800 // variadic templates: either a non-MS compiler or VS >= 2013 -#define MOODYCAMEL_HAS_EMPLACE 1 -#endif -#endif - -#ifdef AE_VCPP -#pragma warning(push) -#pragma warning(disable: 4324) // structure was padded due to __declspec(align()) -#pragma warning(disable: 4820) // padding was added -#pragma warning(disable: 4127) // conditional expression is constant -#endif - -namespace moodycamel { - -template -class ReaderWriterQueue -{ - // Design: Based on a queue-of-queues. The low-level queues are just - // circular buffers with front and tail indices indicating where the - // next element to dequeue is and where the next element can be enqueued, - // respectively. Each low-level queue is called a "block". Each block - // wastes exactly one element's worth of space to keep the design simple - // (if front == tail then the queue is empty, and can't be full). - // The high-level queue is a circular linked list of blocks; again there - // is a front and tail, but this time they are pointers to the blocks. - // The front block is where the next element to be dequeued is, provided - // the block is not empty. The back block is where elements are to be - // enqueued, provided the block is not full. - // The producer thread owns all the tail indices/pointers. The consumer - // thread owns all the front indices/pointers. Both threads read each - // other's variables, but only the owning thread updates them. E.g. After - // the consumer reads the producer's tail, the tail may change before the - // consumer is done dequeuing an object, but the consumer knows the tail - // will never go backwards, only forwards. - // If there is no room to enqueue an object, an additional block (of - // equal size to the last block) is added. Blocks are never removed. - -public: - typedef T value_type; - - // Constructs a queue that can hold maxSize elements without further - // allocations. If more than MAX_BLOCK_SIZE elements are requested, - // then several blocks of MAX_BLOCK_SIZE each are reserved (including - // at least one extra buffer block). - AE_NO_TSAN explicit ReaderWriterQueue(size_t maxSize = 15) -#ifndef NDEBUG - : enqueuing(false) - ,dequeuing(false) -#endif - { - assert(maxSize > 0); - assert(MAX_BLOCK_SIZE == ceilToPow2(MAX_BLOCK_SIZE) && "MAX_BLOCK_SIZE must be a power of 2"); - assert(MAX_BLOCK_SIZE >= 2 && "MAX_BLOCK_SIZE must be at least 2"); - - Block* firstBlock = nullptr; - - largestBlockSize = ceilToPow2(maxSize + 1); // We need a spare slot to fit maxSize elements in the block - if (largestBlockSize > MAX_BLOCK_SIZE * 2) { - // We need a spare block in case the producer is writing to a different block the consumer is reading from, and - // wants to enqueue the maximum number of elements. We also need a spare element in each block to avoid the ambiguity - // between front == tail meaning "empty" and "full". - // So the effective number of slots that are guaranteed to be usable at any time is the block size - 1 times the - // number of blocks - 1. Solving for maxSize and applying a ceiling to the division gives us (after simplifying): - size_t initialBlockCount = (maxSize + MAX_BLOCK_SIZE * 2 - 3) / (MAX_BLOCK_SIZE - 1); - largestBlockSize = MAX_BLOCK_SIZE; - Block* lastBlock = nullptr; - for (size_t i = 0; i != initialBlockCount; ++i) { - auto block = make_block(largestBlockSize); - if (block == nullptr) { -#ifdef MOODYCAMEL_EXCEPTIONS_ENABLED - throw std::bad_alloc(); -#else - abort(); -#endif - } - if (firstBlock == nullptr) { - firstBlock = block; - } - else { - lastBlock->next = block; - } - lastBlock = block; - block->next = firstBlock; - } - } - else { - firstBlock = make_block(largestBlockSize); - if (firstBlock == nullptr) { -#ifdef MOODYCAMEL_EXCEPTIONS_ENABLED - throw std::bad_alloc(); -#else - abort(); -#endif - } - firstBlock->next = firstBlock; - } - frontBlock = firstBlock; - tailBlock = firstBlock; - - // Make sure the reader/writer threads will have the initialized memory setup above: - fence(memory_order_sync); - } - - // Note: The queue should not be accessed concurrently while it's - // being moved. It's up to the user to synchronize this. - AE_NO_TSAN ReaderWriterQueue(ReaderWriterQueue&& other) - : frontBlock(other.frontBlock.load()), - tailBlock(other.tailBlock.load()), - largestBlockSize(other.largestBlockSize) -#ifndef NDEBUG - ,enqueuing(false) - ,dequeuing(false) -#endif - { - other.largestBlockSize = 32; - Block* b = other.make_block(other.largestBlockSize); - if (b == nullptr) { -#ifdef MOODYCAMEL_EXCEPTIONS_ENABLED - throw std::bad_alloc(); -#else - abort(); -#endif - } - b->next = b; - other.frontBlock = b; - other.tailBlock = b; - } - - // Note: The queue should not be accessed concurrently while it's - // being moved. It's up to the user to synchronize this. - ReaderWriterQueue& operator=(ReaderWriterQueue&& other) AE_NO_TSAN - { - Block* b = frontBlock.load(); - frontBlock = other.frontBlock.load(); - other.frontBlock = b; - b = tailBlock.load(); - tailBlock = other.tailBlock.load(); - other.tailBlock = b; - std::swap(largestBlockSize, other.largestBlockSize); - return *this; - } - - // Note: The queue should not be accessed concurrently while it's - // being deleted. It's up to the user to synchronize this. - AE_NO_TSAN ~ReaderWriterQueue() - { - // Make sure we get the latest version of all variables from other CPUs: - fence(memory_order_sync); - - // Destroy any remaining objects in queue and free memory - Block* frontBlock_ = frontBlock; - Block* block = frontBlock_; - do { - Block* nextBlock = block->next; - size_t blockFront = block->front; - size_t blockTail = block->tail; - - for (size_t i = blockFront; i != blockTail; i = (i + 1) & block->sizeMask) { - auto element = reinterpret_cast(block->data + i * sizeof(T)); - element->~T(); - (void)element; - } - - auto rawBlock = block->rawThis; - block->~Block(); - std::free(rawBlock); - block = nextBlock; - } while (block != frontBlock_); - } - - - // Enqueues a copy of element if there is room in the queue. - // Returns true if the element was enqueued, false otherwise. - // Does not allocate memory. - AE_FORCEINLINE bool try_enqueue(T const& element) AE_NO_TSAN - { - return inner_enqueue(element); - } - - // Enqueues a moved copy of element if there is room in the queue. - // Returns true if the element was enqueued, false otherwise. - // Does not allocate memory. - AE_FORCEINLINE bool try_enqueue(T&& element) AE_NO_TSAN - { - return inner_enqueue(std::forward(element)); - } - -#if MOODYCAMEL_HAS_EMPLACE - // Like try_enqueue() but with emplace semantics (i.e. construct-in-place). - template - AE_FORCEINLINE bool try_emplace(Args&&... args) AE_NO_TSAN - { - return inner_enqueue(std::forward(args)...); - } -#endif - - // Enqueues a copy of element on the queue. - // Allocates an additional block of memory if needed. - // Only fails (returns false) if memory allocation fails. - AE_FORCEINLINE bool enqueue(T const& element) AE_NO_TSAN - { - return inner_enqueue(element); - } - - // Enqueues a moved copy of element on the queue. - // Allocates an additional block of memory if needed. - // Only fails (returns false) if memory allocation fails. - AE_FORCEINLINE bool enqueue(T&& element) AE_NO_TSAN - { - return inner_enqueue(std::forward(element)); - } - -#if MOODYCAMEL_HAS_EMPLACE - // Like enqueue() but with emplace semantics (i.e. construct-in-place). - template - AE_FORCEINLINE bool emplace(Args&&... args) AE_NO_TSAN - { - return inner_enqueue(std::forward(args)...); - } -#endif - - // Attempts to dequeue an element; if the queue is empty, - // returns false instead. If the queue has at least one element, - // moves front to result using operator=, then returns true. - template - bool try_dequeue(U& result) AE_NO_TSAN - { -#ifndef NDEBUG - ReentrantGuard guard(this->dequeuing); -#endif - - // High-level pseudocode: - // Remember where the tail block is - // If the front block has an element in it, dequeue it - // Else - // If front block was the tail block when we entered the function, return false - // Else advance to next block and dequeue the item there - - // Note that we have to use the value of the tail block from before we check if the front - // block is full or not, in case the front block is empty and then, before we check if the - // tail block is at the front block or not, the producer fills up the front block *and - // moves on*, which would make us skip a filled block. Seems unlikely, but was consistently - // reproducible in practice. - // In order to avoid overhead in the common case, though, we do a double-checked pattern - // where we have the fast path if the front block is not empty, then read the tail block, - // then re-read the front block and check if it's not empty again, then check if the tail - // block has advanced. - - Block* frontBlock_ = frontBlock.load(); - size_t blockTail = frontBlock_->localTail; - size_t blockFront = frontBlock_->front.load(); - - if (blockFront != blockTail || blockFront != (frontBlock_->localTail = frontBlock_->tail.load())) { - fence(memory_order_acquire); - - non_empty_front_block: - // Front block not empty, dequeue from here - auto element = reinterpret_cast(frontBlock_->data + blockFront * sizeof(T)); - result = std::move(*element); - element->~T(); - - blockFront = (blockFront + 1) & frontBlock_->sizeMask; - - fence(memory_order_release); - frontBlock_->front = blockFront; - } - else if (frontBlock_ != tailBlock.load()) { - fence(memory_order_acquire); - - frontBlock_ = frontBlock.load(); - blockTail = frontBlock_->localTail = frontBlock_->tail.load(); - blockFront = frontBlock_->front.load(); - fence(memory_order_acquire); - - if (blockFront != blockTail) { - // Oh look, the front block isn't empty after all - goto non_empty_front_block; - } - - // Front block is empty but there's another block ahead, advance to it - Block* nextBlock = frontBlock_->next; - // Don't need an acquire fence here since next can only ever be set on the tailBlock, - // and we're not the tailBlock, and we did an acquire earlier after reading tailBlock which - // ensures next is up-to-date on this CPU in case we recently were at tailBlock. - - size_t nextBlockFront = nextBlock->front.load(); - size_t nextBlockTail = nextBlock->localTail = nextBlock->tail.load(); - fence(memory_order_acquire); - - // Since the tailBlock is only ever advanced after being written to, - // we know there's for sure an element to dequeue on it - assert(nextBlockFront != nextBlockTail); - AE_UNUSED(nextBlockTail); - - // We're done with this block, let the producer use it if it needs - fence(memory_order_release); // Expose possibly pending changes to frontBlock->front from last dequeue - frontBlock = frontBlock_ = nextBlock; - - compiler_fence(memory_order_release); // Not strictly needed - - auto element = reinterpret_cast(frontBlock_->data + nextBlockFront * sizeof(T)); - - result = std::move(*element); - element->~T(); - - nextBlockFront = (nextBlockFront + 1) & frontBlock_->sizeMask; - - fence(memory_order_release); - frontBlock_->front = nextBlockFront; - } - else { - // No elements in current block and no other block to advance to - return false; - } - - return true; - } - - - // Returns a pointer to the front element in the queue (the one that - // would be removed next by a call to `try_dequeue` or `pop`). If the - // queue appears empty at the time the method is called, nullptr is - // returned instead. - // Must be called only from the consumer thread. - T* peek() AE_NO_TSAN - { -#ifndef NDEBUG - ReentrantGuard guard(this->dequeuing); -#endif - // See try_dequeue() for reasoning - - Block* frontBlock_ = frontBlock.load(); - size_t blockTail = frontBlock_->localTail; - size_t blockFront = frontBlock_->front.load(); - - if (blockFront != blockTail || blockFront != (frontBlock_->localTail = frontBlock_->tail.load())) { - fence(memory_order_acquire); - non_empty_front_block: - return reinterpret_cast(frontBlock_->data + blockFront * sizeof(T)); - } - else if (frontBlock_ != tailBlock.load()) { - fence(memory_order_acquire); - frontBlock_ = frontBlock.load(); - blockTail = frontBlock_->localTail = frontBlock_->tail.load(); - blockFront = frontBlock_->front.load(); - fence(memory_order_acquire); - - if (blockFront != blockTail) { - goto non_empty_front_block; - } - - Block* nextBlock = frontBlock_->next; - - size_t nextBlockFront = nextBlock->front.load(); - fence(memory_order_acquire); - - assert(nextBlockFront != nextBlock->tail.load()); - return reinterpret_cast(nextBlock->data + nextBlockFront * sizeof(T)); - } - - return nullptr; - } - - // Removes the front element from the queue, if any, without returning it. - // Returns true on success, or false if the queue appeared empty at the time - // `pop` was called. - bool pop() AE_NO_TSAN - { -#ifndef NDEBUG - ReentrantGuard guard(this->dequeuing); -#endif - // See try_dequeue() for reasoning - - Block* frontBlock_ = frontBlock.load(); - size_t blockTail = frontBlock_->localTail; - size_t blockFront = frontBlock_->front.load(); - - if (blockFront != blockTail || blockFront != (frontBlock_->localTail = frontBlock_->tail.load())) { - fence(memory_order_acquire); - - non_empty_front_block: - auto element = reinterpret_cast(frontBlock_->data + blockFront * sizeof(T)); - element->~T(); - - blockFront = (blockFront + 1) & frontBlock_->sizeMask; - - fence(memory_order_release); - frontBlock_->front = blockFront; - } - else if (frontBlock_ != tailBlock.load()) { - fence(memory_order_acquire); - frontBlock_ = frontBlock.load(); - blockTail = frontBlock_->localTail = frontBlock_->tail.load(); - blockFront = frontBlock_->front.load(); - fence(memory_order_acquire); - - if (blockFront != blockTail) { - goto non_empty_front_block; - } - - // Front block is empty but there's another block ahead, advance to it - Block* nextBlock = frontBlock_->next; - - size_t nextBlockFront = nextBlock->front.load(); - size_t nextBlockTail = nextBlock->localTail = nextBlock->tail.load(); - fence(memory_order_acquire); - - assert(nextBlockFront != nextBlockTail); - AE_UNUSED(nextBlockTail); - - fence(memory_order_release); - frontBlock = frontBlock_ = nextBlock; - - compiler_fence(memory_order_release); - - auto element = reinterpret_cast(frontBlock_->data + nextBlockFront * sizeof(T)); - element->~T(); - - nextBlockFront = (nextBlockFront + 1) & frontBlock_->sizeMask; - - fence(memory_order_release); - frontBlock_->front = nextBlockFront; - } - else { - // No elements in current block and no other block to advance to - return false; - } - - return true; - } - - // Returns the approximate number of items currently in the queue. - // Safe to call from both the producer and consumer threads. - inline size_t size_approx() const AE_NO_TSAN - { - size_t result = 0; - Block* frontBlock_ = frontBlock.load(); - Block* block = frontBlock_; - do { - fence(memory_order_acquire); - size_t blockFront = block->front.load(); - size_t blockTail = block->tail.load(); - result += (blockTail - blockFront) & block->sizeMask; - block = block->next.load(); - } while (block != frontBlock_); - return result; - } - - -private: - enum AllocationMode { CanAlloc, CannotAlloc }; - -#if MOODYCAMEL_HAS_EMPLACE - template - bool inner_enqueue(Args&&... args) AE_NO_TSAN -#else - template - bool inner_enqueue(U&& element) AE_NO_TSAN -#endif - { -#ifndef NDEBUG - ReentrantGuard guard(this->enqueuing); -#endif - - // High-level pseudocode (assuming we're allowed to alloc a new block): - // If room in tail block, add to tail - // Else check next block - // If next block is not the head block, enqueue on next block - // Else create a new block and enqueue there - // Advance tail to the block we just enqueued to - - Block* tailBlock_ = tailBlock.load(); - size_t blockFront = tailBlock_->localFront; - size_t blockTail = tailBlock_->tail.load(); - - size_t nextBlockTail = (blockTail + 1) & tailBlock_->sizeMask; - if (nextBlockTail != blockFront || nextBlockTail != (tailBlock_->localFront = tailBlock_->front.load())) { - fence(memory_order_acquire); - // This block has room for at least one more element - char* location = tailBlock_->data + blockTail * sizeof(T); -#if MOODYCAMEL_HAS_EMPLACE - new (location) T(std::forward(args)...); -#else - new (location) T(std::forward(element)); -#endif - - fence(memory_order_release); - tailBlock_->tail = nextBlockTail; - } - else { - fence(memory_order_acquire); - if (tailBlock_->next.load() != frontBlock) { - // Note that the reason we can't advance to the frontBlock and start adding new entries there - // is because if we did, then dequeue would stay in that block, eventually reading the new values, - // instead of advancing to the next full block (whose values were enqueued first and so should be - // consumed first). - - fence(memory_order_acquire); // Ensure we get latest writes if we got the latest frontBlock - - // tailBlock is full, but there's a free block ahead, use it - Block* tailBlockNext = tailBlock_->next.load(); - size_t nextBlockFront = tailBlockNext->localFront = tailBlockNext->front.load(); - nextBlockTail = tailBlockNext->tail.load(); - fence(memory_order_acquire); - - // This block must be empty since it's not the head block and we - // go through the blocks in a circle - assert(nextBlockFront == nextBlockTail); - tailBlockNext->localFront = nextBlockFront; - - char* location = tailBlockNext->data + nextBlockTail * sizeof(T); -#if MOODYCAMEL_HAS_EMPLACE - new (location) T(std::forward(args)...); -#else - new (location) T(std::forward(element)); -#endif - - tailBlockNext->tail = (nextBlockTail + 1) & tailBlockNext->sizeMask; - - fence(memory_order_release); - tailBlock = tailBlockNext; - } - else if (canAlloc == CanAlloc) { - // tailBlock is full and there's no free block ahead; create a new block - auto newBlockSize = largestBlockSize >= MAX_BLOCK_SIZE ? largestBlockSize : largestBlockSize * 2; - auto newBlock = make_block(newBlockSize); - if (newBlock == nullptr) { - // Could not allocate a block! - return false; - } - largestBlockSize = newBlockSize; - -#if MOODYCAMEL_HAS_EMPLACE - new (newBlock->data) T(std::forward(args)...); -#else - new (newBlock->data) T(std::forward(element)); -#endif - assert(newBlock->front == 0); - newBlock->tail = newBlock->localTail = 1; - - newBlock->next = tailBlock_->next.load(); - tailBlock_->next = newBlock; - - // Might be possible for the dequeue thread to see the new tailBlock->next - // *without* seeing the new tailBlock value, but this is OK since it can't - // advance to the next block until tailBlock is set anyway (because the only - // case where it could try to read the next is if it's already at the tailBlock, - // and it won't advance past tailBlock in any circumstance). - - fence(memory_order_release); - tailBlock = newBlock; - } - else if (canAlloc == CannotAlloc) { - // Would have had to allocate a new block to enqueue, but not allowed - return false; - } - else { - assert(false && "Should be unreachable code"); - return false; - } - } - - return true; - } - - - // Disable copying - ReaderWriterQueue(ReaderWriterQueue const&) { } - - // Disable assignment - ReaderWriterQueue& operator=(ReaderWriterQueue const&) { } - - - - AE_FORCEINLINE static size_t ceilToPow2(size_t x) - { - // From http://graphics.stanford.edu/~seander/bithacks.html#RoundUpPowerOf2 - --x; - x |= x >> 1; - x |= x >> 2; - x |= x >> 4; - for (size_t i = 1; i < sizeof(size_t); i <<= 1) { - x |= x >> (i << 3); - } - ++x; - return x; - } - - template - static AE_FORCEINLINE char* align_for(char* ptr) AE_NO_TSAN - { - const std::size_t alignment = std::alignment_of::value; - return ptr + (alignment - (reinterpret_cast(ptr) % alignment)) % alignment; - } -private: -#ifndef NDEBUG - struct ReentrantGuard - { - AE_NO_TSAN ReentrantGuard(bool& _inSection) - : inSection(_inSection) - { - assert(!inSection && "Concurrent (or re-entrant) enqueue or dequeue operation detected (only one thread at a time may hold the producer or consumer role)"); - inSection = true; - } - - AE_NO_TSAN ~ReentrantGuard() { inSection = false; } - - private: - ReentrantGuard& operator=(ReentrantGuard const&); - - private: - bool& inSection; - }; -#endif - - struct Block - { - // Avoid false-sharing by putting highly contended variables on their own cache lines - weak_atomic front; // (Atomic) Elements are read from here - size_t localTail; // An uncontended shadow copy of tail, owned by the consumer - - char cachelineFiller0[MOODYCAMEL_CACHE_LINE_SIZE - sizeof(weak_atomic) - sizeof(size_t)]; - weak_atomic tail; // (Atomic) Elements are enqueued here - size_t localFront; - - char cachelineFiller1[MOODYCAMEL_CACHE_LINE_SIZE - sizeof(weak_atomic) - sizeof(size_t)]; // next isn't very contended, but we don't want it on the same cache line as tail (which is) - weak_atomic next; // (Atomic) - - char* data; // Contents (on heap) are aligned to T's alignment - - const size_t sizeMask; - - - // size must be a power of two (and greater than 0) - AE_NO_TSAN Block(size_t const& _size, char* _rawThis, char* _data) - : front(0), localTail(0), tail(0), localFront(0), next(nullptr), data(_data), sizeMask(_size - 1), rawThis(_rawThis) - { - } - - private: - // C4512 - Assignment operator could not be generated - Block& operator=(Block const&); - - public: - char* rawThis; - }; - - - static Block* make_block(size_t capacity) AE_NO_TSAN - { - // Allocate enough memory for the block itself, as well as all the elements it will contain - auto size = sizeof(Block) + std::alignment_of::value - 1; - size += sizeof(T) * capacity + std::alignment_of::value - 1; - auto newBlockRaw = static_cast(std::malloc(size)); - if (newBlockRaw == nullptr) { - return nullptr; - } - - auto newBlockAligned = align_for(newBlockRaw); - auto newBlockData = align_for(newBlockAligned + sizeof(Block)); - return new (newBlockAligned) Block(capacity, newBlockRaw, newBlockData); - } - -private: - weak_atomic frontBlock; // (Atomic) Elements are enqueued to this block - - char cachelineFiller[MOODYCAMEL_CACHE_LINE_SIZE - sizeof(weak_atomic)]; - weak_atomic tailBlock; // (Atomic) Elements are dequeued from this block - - size_t largestBlockSize; - -#ifndef NDEBUG - bool enqueuing; - bool dequeuing; -#endif -}; - -// Like ReaderWriterQueue, but also providees blocking operations -template -class BlockingReaderWriterQueue -{ -private: - typedef ::moodycamel::ReaderWriterQueue ReaderWriterQueue; - -public: - explicit BlockingReaderWriterQueue(size_t maxSize = 15) AE_NO_TSAN - : inner(maxSize), sema(new spsc_sema::LightweightSemaphore()) - { } - - BlockingReaderWriterQueue(BlockingReaderWriterQueue&& other) AE_NO_TSAN - : inner(std::move(other.inner)), sema(std::move(other.sema)) - { } - - BlockingReaderWriterQueue& operator=(BlockingReaderWriterQueue&& other) AE_NO_TSAN - { - std::swap(sema, other.sema); - std::swap(inner, other.inner); - return *this; - } - - - // Enqueues a copy of element if there is room in the queue. - // Returns true if the element was enqueued, false otherwise. - // Does not allocate memory. - AE_FORCEINLINE bool try_enqueue(T const& element) AE_NO_TSAN - { - if (inner.try_enqueue(element)) { - sema->signal(); - return true; - } - return false; - } - - // Enqueues a moved copy of element if there is room in the queue. - // Returns true if the element was enqueued, false otherwise. - // Does not allocate memory. - AE_FORCEINLINE bool try_enqueue(T&& element) AE_NO_TSAN - { - if (inner.try_enqueue(std::forward(element))) { - sema->signal(); - return true; - } - return false; - } - - - // Enqueues a copy of element on the queue. - // Allocates an additional block of memory if needed. - // Only fails (returns false) if memory allocation fails. - AE_FORCEINLINE bool enqueue(T const& element) AE_NO_TSAN - { - if (inner.enqueue(element)) { - sema->signal(); - return true; - } - return false; - } - - // Enqueues a moved copy of element on the queue. - // Allocates an additional block of memory if needed. - // Only fails (returns false) if memory allocation fails. - AE_FORCEINLINE bool enqueue(T&& element) AE_NO_TSAN - { - if (inner.enqueue(std::forward(element))) { - sema->signal(); - return true; - } - return false; - } - - - // Attempts to dequeue an element; if the queue is empty, - // returns false instead. If the queue has at least one element, - // moves front to result using operator=, then returns true. - template - bool try_dequeue(U& result) AE_NO_TSAN - { - if (sema->tryWait()) { - bool success = inner.try_dequeue(result); - assert(success); - AE_UNUSED(success); - return true; - } - return false; - } - - - // Attempts to dequeue an element; if the queue is empty, - // waits until an element is available, then dequeues it. - template - void wait_dequeue(U& result) AE_NO_TSAN - { - sema->wait(); - bool success = inner.try_dequeue(result); - AE_UNUSED(result); - assert(success); - AE_UNUSED(success); - } - - - // Attempts to dequeue an element; if the queue is empty, - // waits until an element is available up to the specified timeout, - // then dequeues it and returns true, or returns false if the timeout - // expires before an element can be dequeued. - // Using a negative timeout indicates an indefinite timeout, - // and is thus functionally equivalent to calling wait_dequeue. - template - bool wait_dequeue_timed(U& result, std::int64_t timeout_usecs) AE_NO_TSAN - { - if (!sema->wait(timeout_usecs)) { - return false; - } - bool success = inner.try_dequeue(result); - AE_UNUSED(result); - assert(success); - AE_UNUSED(success); - return true; - } - - -#if __cplusplus > 199711L || _MSC_VER >= 1700 - // Attempts to dequeue an element; if the queue is empty, - // waits until an element is available up to the specified timeout, - // then dequeues it and returns true, or returns false if the timeout - // expires before an element can be dequeued. - // Using a negative timeout indicates an indefinite timeout, - // and is thus functionally equivalent to calling wait_dequeue. - template - inline bool wait_dequeue_timed(U& result, std::chrono::duration const& timeout) AE_NO_TSAN - { - return wait_dequeue_timed(result, std::chrono::duration_cast(timeout).count()); - } -#endif - - - // Returns a pointer to the front element in the queue (the one that - // would be removed next by a call to `try_dequeue` or `pop`). If the - // queue appears empty at the time the method is called, nullptr is - // returned instead. - // Must be called only from the consumer thread. - AE_FORCEINLINE T* peek() AE_NO_TSAN - { - return inner.peek(); - } - - // Removes the front element from the queue, if any, without returning it. - // Returns true on success, or false if the queue appeared empty at the time - // `pop` was called. - AE_FORCEINLINE bool pop() AE_NO_TSAN - { - if (sema->tryWait()) { - bool result = inner.pop(); - assert(result); - AE_UNUSED(result); - return true; - } - return false; - } - - // Returns the approximate number of items currently in the queue. - // Safe to call from both the producer and consumer threads. - AE_FORCEINLINE size_t size_approx() const AE_NO_TSAN - { - return sema->availableApprox(); - } - - -private: - // Disable copying & assignment - BlockingReaderWriterQueue(BlockingReaderWriterQueue const&) { } - BlockingReaderWriterQueue& operator=(BlockingReaderWriterQueue const&) { } - -private: - ReaderWriterQueue inner; - std::unique_ptr sema; -}; - -} // end namespace moodycamel - -#ifdef AE_VCPP -#pragma warning(pop) -#endif diff --git a/include/zmq.h b/include/zmq.h index edf28efd2b..d0174c5a1a 100644 --- a/include/zmq.h +++ b/include/zmq.h @@ -268,6 +268,8 @@ typedef void(zmq_free_fn) (void *data_, void *hint_); ZMQ_EXPORT int zmq_msg_init (zmq_msg_t *msg_); ZMQ_EXPORT int zmq_msg_init_size (zmq_msg_t *msg_, size_t size_); +ZMQ_EXPORT int +zmq_msg_init_allocator (zmq_msg_t *msg_, size_t size_, void *allocator_); ZMQ_EXPORT int zmq_msg_init_data ( zmq_msg_t *msg_, void *data_, size_t size_, zmq_free_fn *ffn_, void *hint_); ZMQ_EXPORT int zmq_msg_send (zmq_msg_t *msg_, void *s_, int flags_); @@ -669,6 +671,7 @@ ZMQ_EXPORT void zmq_threadclose (void *thread_); /* DRAFT Context options */ #define ZMQ_ZERO_COPY_RECV 10 +//#define ZMQ_MSG_ALLOCATOR 11 /* DRAFT Context methods. */ ZMQ_EXPORT int zmq_ctx_set_ext (void *context_, @@ -680,6 +683,17 @@ ZMQ_EXPORT int zmq_ctx_get_ext (void *context_, void *optval_, size_t *optvallen_); +/* ZMQ-provided message-pool implementations. */ +// default allocator using malloc/free +#define ZMQ_MSG_ALLOCATOR_DEFAULT 0 +// using internally a SPSC queue (cannot be used with inproc maybe?) or perhaps an MPMC queue anyway +#define ZMQ_MSG_ALLOCATOR_PER_THREAD_POOL 1 +// using internally a MPMC queue +#define ZMQ_MSG_ALLOCATOR_GLOBAL_POOL 2 + +ZMQ_EXPORT void *zmq_msg_allocator_new (int type_); +ZMQ_EXPORT int zmq_msg_allocator_destroy (void **allocator_); + /* DRAFT Socket methods. */ ZMQ_EXPORT int zmq_join (void *s, const char *group); ZMQ_EXPORT int zmq_leave (void *s, const char *group); diff --git a/perf/remote_thr.cpp b/perf/remote_thr.cpp index 6969e8c25f..3f47234622 100644 --- a/perf/remote_thr.cpp +++ b/perf/remote_thr.cpp @@ -27,97 +27,17 @@ along with this program. If not, see . */ +#include "../src/platform.hpp" #include "../include/zmq.h" #include #include #include -#include - -#include "readerwriterqueue.h" // keys are arbitrary but must match local_lat.cpp const char server_pubkey[] = "DX4nh=yUn{-9ugra0X3Src4SU-4xTgqxcYY.+raw_data); - - return true; - } - - static void - deallocate_msg (void *data_, - void *hint_) // producer thread: ZMQ background IO thread - { - ZmqMessagePool *pPool = reinterpret_cast (hint_); - - // recover the beginning of this msg_block: - uint8_t *data_ptr_ = (uint8_t *) data_; - msg_block_t *to_return = - (msg_block_t *) (data_ptr_ - SIZE_OF_CONTENT_T_USED_BY_ZMQ_VLM); - assert (to_return->canary == 0xAB); - - // produce a new free msg block: - pPool->m_free_list.enqueue (to_return); - } - - size_t size () const { return m_free_list.size_approx (); } - - private: - msg_block_t m_storage[MAX_ACTIVE_MESSAGES]; - moodycamel::ReaderWriterQueue m_free_list; -}; - - int main (int argc, char *argv[]) { const char *connect_to; @@ -148,6 +68,11 @@ int main (int argc, char *argv[]) return -1; } +#ifdef ZMQ_BUILD_DRAFT_API + // EXPERIMENTAL ALLOCATOR FOR MSG_T + void *allocator = zmq_msg_allocator_new (ZMQ_MSG_ALLOCATOR_GLOBAL_POOL); +#endif + s = zmq_socket (ctx, ZMQ_PUSH); if (!s) { printf ("error in zmq_socket: %s\n", zmq_strerror (errno)); @@ -185,9 +110,12 @@ int main (int argc, char *argv[]) return -1; } -#if 0 for (i = 0; i != message_count; i++) { +#ifdef ZMQ_BUILD_DRAFT_API + rc = zmq_msg_init_allocator (&msg, message_size, allocator); +#else rc = zmq_msg_init_size (&msg, message_size); +#endif if (rc != 0) { printf ("error in zmq_msg_init_size: %s\n", zmq_strerror (errno)); return -1; @@ -203,32 +131,6 @@ int main (int argc, char *argv[]) return -1; } } -#else - printf ("msg block size: %zu; max msg size: %d\n", sizeof (msg_block_t), - MAX_MESSAGE_SIZE); - ZmqMessagePool pool; - for (i = 0; i != message_count; i++) { - pool.allocate_msg (&msg, message_size); - - // to be fair when comparing the results generated by the other #if/#endif branch - // avoid any kind of initialization of message memory: - //memset (zmq_msg_data (&msg), message_size, 0xAB); - - rc = zmq_sendmsg (s, &msg, 0); - if (rc < 0) { - printf ("error in zmq_sendmsg: %s\n", zmq_strerror (errno)); - return -1; - } - rc = zmq_msg_close (&msg); - if (rc != 0) { - printf ("error in zmq_msg_close: %s\n", zmq_strerror (errno)); - return -1; - } - - //if ((i % 1000) == 0) - // printf ("mempool msg size: %zu\n", pool.size ()); - } -#endif rc = zmq_close (s); if (rc != 0) { @@ -242,5 +144,11 @@ int main (int argc, char *argv[]) return -1; } +#ifdef ZMQ_BUILD_DRAFT_API + // IMPORTANT: destroy the allocator only after zmq_ctx_term() since otherwise + // some zmq_msg_t may still be "in fly" + zmq_msg_allocator_destroy (&allocator); +#endif + return 0; } diff --git a/src/allocator.cpp b/src/allocator.cpp new file mode 100644 index 0000000000..ff6b6320fa --- /dev/null +++ b/src/allocator.cpp @@ -0,0 +1,97 @@ +/* + Copyright (c) 2007-2016 Contributors as noted in the AUTHORS file + + This file is part of libzmq, the ZeroMQ core engine in C++. + + libzmq is free software; you can redistribute it and/or modify it under + the terms of the GNU Lesser General Public License (LGPL) as published + by the Free Software Foundation; either version 3 of the License, or + (at your option) any later version. + + As a special exception, the Contributors give you permission to link + this library with independent modules to produce an executable, + regardless of the license terms of these independent modules, and to + copy and distribute the resulting executable under terms of your choice, + provided that you also meet, for each linked independent module, the + terms and conditions of the license of that module. An independent + module is a module which is not derived from or based on this library. + If you modify this library, you must extend this exception to your + version of the library. + + libzmq is distributed in the hope that it will be useful, but WITHOUT + ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or + FITNESS FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public + License for more details. + + You should have received a copy of the GNU Lesser General Public License + along with this program. If not, see . +*/ + +#include "precompiled.hpp" +#include "allocator.hpp" + + +zmq::allocator_t::allocator_t () +{ + _type = ZMQ_MSG_ALLOCATOR_DEFAULT; + _tag = 0xCAFEEBEB; +} + +size_t zmq::allocator_t::size () const +{ + switch (_type) { + case ZMQ_MSG_ALLOCATOR_DEFAULT: + return 0; + + // using internally a SPSC queue (cannot be used with inproc maybe?) or perhaps an MPMC queue anyway + case ZMQ_MSG_ALLOCATOR_PER_THREAD_POOL: + return 0; + + // using internally a MPMC queue + case ZMQ_MSG_ALLOCATOR_GLOBAL_POOL: + return _global_pool.size (); + + default: + return 0; + } +} + + +void *zmq::allocator_t::allocate (size_t len) +{ + switch (_type) { + case ZMQ_MSG_ALLOCATOR_DEFAULT: + return malloc (len); + + // using internally a SPSC queue (cannot be used with inproc maybe?) or perhaps an MPMC queue anyway + case ZMQ_MSG_ALLOCATOR_PER_THREAD_POOL: + // FIXME + return NULL; + + // using internally a MPMC queue + case ZMQ_MSG_ALLOCATOR_GLOBAL_POOL: + return _global_pool.allocate_msg (len); + } + return NULL; +} + +void zmq::allocator_t::deallocate_msg (void *data_, void *hint_) +{ + allocator_t *alloc = reinterpret_cast (hint_); + switch (alloc->_type) { + case ZMQ_MSG_ALLOCATOR_DEFAULT: + free (data_); + return; + + // using internally a SPSC queue (cannot be used with inproc maybe?) or perhaps an MPMC queue anyway + case ZMQ_MSG_ALLOCATOR_PER_THREAD_POOL: + // FIXME + return; + + // using internally a MPMC queue + case ZMQ_MSG_ALLOCATOR_GLOBAL_POOL: + zmq::msg_t::content_t *msg_content = + (zmq::msg_t::content_t *) data_; + alloc->_global_pool.deallocate_msg (msg_content, msg_content->size); + } +} diff --git a/src/allocator.hpp b/src/allocator.hpp new file mode 100644 index 0000000000..8cac7e8584 --- /dev/null +++ b/src/allocator.hpp @@ -0,0 +1,181 @@ +/* + Copyright (c) 2007-2016 Contributors as noted in the AUTHORS file + + This file is part of libzmq, the ZeroMQ core engine in C++. + + libzmq is free software; you can redistribute it and/or modify it under + the terms of the GNU Lesser General Public License (LGPL) as published + by the Free Software Foundation; either version 3 of the License, or + (at your option) any later version. + + As a special exception, the Contributors give you permission to link + this library with independent modules to produce an executable, + regardless of the license terms of these independent modules, and to + copy and distribute the resulting executable under terms of your choice, + provided that you also meet, for each linked independent module, the + terms and conditions of the license of that module. An independent + module is a module which is not derived from or based on this library. + If you modify this library, you must extend this exception to your + version of the library. + + libzmq is distributed in the hope that it will be useful, but WITHOUT + ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or + FITNESS FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public + License for more details. + + You should have received a copy of the GNU Lesser General Public License + along with this program. If not, see . +*/ + +#ifndef __ZMQ_MEMORYPOOL_HPP_INCLUDED__ +#define __ZMQ_MEMORYPOOL_HPP_INCLUDED__ + +#include +#include "msg.hpp" +#include "concurrentqueue.h" + +// FIXME: we need to grow dynamically the mempool +#define MAX_ACTIVE_MESSAGES (8192) + +namespace zmq +{ +class global_memory_pool_t +{ + typedef struct + { + size_t num_msgs; + // actual user data + uint8_t *raw_data; + } msg_block_t; + + typedef enum + { + MsgBlock_SizeClass_256 = 0, // for messages up to 256B long + MsgBlock_SizeClass_512, + MsgBlock_SizeClass_1024, + MsgBlock_SizeClass_2048, + MsgBlock_SizeClass_4096, + MsgBlock_SizeClass_8192, + + MsgBlock_NumSizeClasses + } MsgBlock_e; + + inline size_t MsgBlockToBytes (MsgBlock_e block_class) + { + switch (block_class) { + case MsgBlock_SizeClass_256: + return 256; + case MsgBlock_SizeClass_512: + return 512; + case MsgBlock_SizeClass_1024: + return 1024; + case MsgBlock_SizeClass_2048: + return 2048; + case MsgBlock_SizeClass_4096: + return 4096; + case MsgBlock_SizeClass_8192: + return 8192; + default: + return 0; + } + } + inline MsgBlock_e BytesToMsgBlock (size_t n) + { + if (n < 256) + return MsgBlock_SizeClass_256; + else if (n < 512) + return MsgBlock_SizeClass_512; + + return MsgBlock_NumSizeClasses; + } + + public: + global_memory_pool_t () + { + // enqueue all available blocks in the free list: + for (int i = 0; i < MsgBlock_NumSizeClasses; i++) { + size_t msg_size = MsgBlockToBytes ((MsgBlock_e) i); + + m_storage[i].num_msgs = MAX_ACTIVE_MESSAGES; + m_storage[i].raw_data = + (uint8_t *) malloc (MAX_ACTIVE_MESSAGES * msg_size); + + uint8_t *msg_memory = m_storage[i].raw_data; + for (int j = 0; j < MAX_ACTIVE_MESSAGES; j++) { + m_free_list[i].enqueue (msg_memory); + msg_memory += msg_size; + } + } + } + ~global_memory_pool_t () {} + + void *allocate_msg (size_t len) // consumer thread: user app thread + { + MsgBlock_e bl = BytesToMsgBlock (len); + assert (bl != MsgBlock_NumSizeClasses); + + // consume 1 block from the list of free msg + uint8_t *next_avail = nullptr; + if (!m_free_list[bl].try_dequeue (next_avail)) { + assert (0); // I want to find out if this ever happens + return NULL; + } + + assert (next_avail); + return next_avail; + } + + void + deallocate_msg (void *data_, + size_t len) // producer thread: ZMQ background IO thread + { + MsgBlock_e bl = BytesToMsgBlock (len); + assert (bl != MsgBlock_NumSizeClasses); + + // produce a new free msg: + m_free_list[bl].enqueue ((uint8_t *) data_); + } + + size_t size () const + { + size_t acc = 0; + for (int i = 0; i < MsgBlock_NumSizeClasses; i++) + acc += m_free_list[i].size_approx (); + return acc; + } + + private: + msg_block_t m_storage[MsgBlock_NumSizeClasses]; + moodycamel::ConcurrentQueue m_free_list[MsgBlock_NumSizeClasses]; +}; + +class allocator_t +{ + public: + allocator_t (); + ~allocator_t () + { + // Mark this instance as dead + _tag = 0xdeadbeef; + } + + void init (int type_) { _type = type_; } + + // allocate() gets called by the consumer thread: the user app thread + void *allocate (size_t len); + + // deallocate_msg() gets called by the producer thread: the ZMQ background IO thread + static void deallocate_msg (void *data_, void *hint_); + + size_t size () const; + bool check_tag () const { return _tag == 0xCAFEEBEB; } + + + private: + int _type; + uint32_t _tag; + global_memory_pool_t _global_pool; +}; +} + +#endif diff --git a/src/concurrentqueue.h b/src/concurrentqueue.h new file mode 100644 index 0000000000..21cb9375aa --- /dev/null +++ b/src/concurrentqueue.h @@ -0,0 +1,3636 @@ +// Provides a C++11 implementation of a multi-producer, multi-consumer lock-free queue. +// An overview, including benchmark results, is provided here: +// http://moodycamel.com/blog/2014/a-fast-general-purpose-lock-free-queue-for-c++ +// The full design is also described in excruciating detail at: +// http://moodycamel.com/blog/2014/detailed-design-of-a-lock-free-queue + +// Simplified BSD license: +// Copyright (c) 2013-2016, Cameron Desrochers. +// All rights reserved. +// +// Redistribution and use in source and binary forms, with or without modification, +// are permitted provided that the following conditions are met: +// +// - Redistributions of source code must retain the above copyright notice, this list of +// conditions and the following disclaimer. +// - Redistributions in binary form must reproduce the above copyright notice, this list of +// conditions and the following disclaimer in the documentation and/or other materials +// provided with the distribution. +// +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY +// EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF +// MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL +// THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, +// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT +// OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) +// HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR +// TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, +// EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + + +#pragma once + +#if defined(__GNUC__) +// Disable -Wconversion warnings (spuriously triggered when Traits::size_t and +// Traits::index_t are set to < 32 bits, causing integer promotion, causing warnings +// upon assigning any computed values) +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wconversion" + +#ifdef MCDBGQ_USE_RELACY +#pragma GCC diagnostic ignored "-Wint-to-pointer-cast" +#endif +#endif + +#if defined(__APPLE__) +#include "TargetConditionals.h" +#endif + +#ifdef MCDBGQ_USE_RELACY +#include "relacy/relacy_std.hpp" +#include "relacy_shims.h" +// We only use malloc/free anyway, and the delete macro messes up `= delete` method declarations. +// We'll override the default trait malloc ourselves without a macro. +#undef new +#undef delete +#undef malloc +#undef free +#else +#include // Requires C++11. Sorry VS2010. +#include +#endif +#include // for max_align_t +#include +#include +#include +#include +#include +#include +#include // for CHAR_BIT +#include +#include // partly for __WINPTHREADS_VERSION if on MinGW-w64 w/ POSIX threading + +// Platform-specific definitions of a numeric thread ID type and an invalid value +namespace moodycamel { namespace details { + template struct thread_id_converter { + typedef thread_id_t thread_id_numeric_size_t; + typedef thread_id_t thread_id_hash_t; + static thread_id_hash_t prehash(thread_id_t const& x) { return x; } + }; +} } +#if defined(MCDBGQ_USE_RELACY) +namespace moodycamel { namespace details { + typedef std::uint32_t thread_id_t; + static const thread_id_t invalid_thread_id = 0xFFFFFFFFU; + static const thread_id_t invalid_thread_id2 = 0xFFFFFFFEU; + static inline thread_id_t thread_id() { return rl::thread_index(); } +} } +#elif defined(_WIN32) || defined(__WINDOWS__) || defined(__WIN32__) +// No sense pulling in windows.h in a header, we'll manually declare the function +// we use and rely on backwards-compatibility for this not to break +extern "C" __declspec(dllimport) unsigned long __stdcall GetCurrentThreadId(void); +namespace moodycamel { namespace details { + static_assert(sizeof(unsigned long) == sizeof(std::uint32_t), "Expected size of unsigned long to be 32 bits on Windows"); + typedef std::uint32_t thread_id_t; + static const thread_id_t invalid_thread_id = 0; // See http://blogs.msdn.com/b/oldnewthing/archive/2004/02/23/78395.aspx + static const thread_id_t invalid_thread_id2 = 0xFFFFFFFFU; // Not technically guaranteed to be invalid, but is never used in practice. Note that all Win32 thread IDs are presently multiples of 4. + static inline thread_id_t thread_id() { return static_cast(::GetCurrentThreadId()); } +} } +#elif defined(__arm__) || defined(_M_ARM) || defined(__aarch64__) || (defined(__APPLE__) && TARGET_OS_IPHONE) +namespace moodycamel { namespace details { + static_assert(sizeof(std::thread::id) == 4 || sizeof(std::thread::id) == 8, "std::thread::id is expected to be either 4 or 8 bytes"); + + typedef std::thread::id thread_id_t; + static const thread_id_t invalid_thread_id; // Default ctor creates invalid ID + + // Note we don't define a invalid_thread_id2 since std::thread::id doesn't have one; it's + // only used if MOODYCAMEL_CPP11_THREAD_LOCAL_SUPPORTED is defined anyway, which it won't + // be. + static inline thread_id_t thread_id() { return std::this_thread::get_id(); } + + template struct thread_id_size { }; + template<> struct thread_id_size<4> { typedef std::uint32_t numeric_t; }; + template<> struct thread_id_size<8> { typedef std::uint64_t numeric_t; }; + + template<> struct thread_id_converter { + typedef thread_id_size::numeric_t thread_id_numeric_size_t; +#ifndef __APPLE__ + typedef std::size_t thread_id_hash_t; +#else + typedef thread_id_numeric_size_t thread_id_hash_t; +#endif + + static thread_id_hash_t prehash(thread_id_t const& x) + { +#ifndef __APPLE__ + return std::hash()(x); +#else + return *reinterpret_cast(&x); +#endif + } + }; +} } +#else +// Use a nice trick from this answer: http://stackoverflow.com/a/8438730/21475 +// In order to get a numeric thread ID in a platform-independent way, we use a thread-local +// static variable's address as a thread identifier :-) +#if defined(__GNUC__) || defined(__INTEL_COMPILER) +#define MOODYCAMEL_THREADLOCAL __thread +#elif defined(_MSC_VER) +#define MOODYCAMEL_THREADLOCAL __declspec(thread) +#else +// Assume C++11 compliant compiler +#define MOODYCAMEL_THREADLOCAL thread_local +#endif +namespace moodycamel { namespace details { + typedef std::uintptr_t thread_id_t; + static const thread_id_t invalid_thread_id = 0; // Address can't be nullptr + static const thread_id_t invalid_thread_id2 = 1; // Member accesses off a null pointer are also generally invalid. Plus it's not aligned. + static inline thread_id_t thread_id() { static MOODYCAMEL_THREADLOCAL int x; return reinterpret_cast(&x); } +} } +#endif + +// Exceptions +#ifndef MOODYCAMEL_EXCEPTIONS_ENABLED +#if (defined(_MSC_VER) && defined(_CPPUNWIND)) || (defined(__GNUC__) && defined(__EXCEPTIONS)) || (!defined(_MSC_VER) && !defined(__GNUC__)) +#define MOODYCAMEL_EXCEPTIONS_ENABLED +#endif +#endif +#ifdef MOODYCAMEL_EXCEPTIONS_ENABLED +#define MOODYCAMEL_TRY try +#define MOODYCAMEL_CATCH(...) catch(__VA_ARGS__) +#define MOODYCAMEL_RETHROW throw +#define MOODYCAMEL_THROW(expr) throw (expr) +#else +#define MOODYCAMEL_TRY if (true) +#define MOODYCAMEL_CATCH(...) else if (false) +#define MOODYCAMEL_RETHROW +#define MOODYCAMEL_THROW(expr) +#endif + +#ifndef MOODYCAMEL_NOEXCEPT +#if !defined(MOODYCAMEL_EXCEPTIONS_ENABLED) +#define MOODYCAMEL_NOEXCEPT +#define MOODYCAMEL_NOEXCEPT_CTOR(type, valueType, expr) true +#define MOODYCAMEL_NOEXCEPT_ASSIGN(type, valueType, expr) true +#elif defined(_MSC_VER) && defined(_NOEXCEPT) && _MSC_VER < 1800 +// VS2012's std::is_nothrow_[move_]constructible is broken and returns true when it shouldn't :-( +// We have to assume *all* non-trivial constructors may throw on VS2012! +#define MOODYCAMEL_NOEXCEPT _NOEXCEPT +#define MOODYCAMEL_NOEXCEPT_CTOR(type, valueType, expr) (std::is_rvalue_reference::value && std::is_move_constructible::value ? std::is_trivially_move_constructible::value : std::is_trivially_copy_constructible::value) +#define MOODYCAMEL_NOEXCEPT_ASSIGN(type, valueType, expr) ((std::is_rvalue_reference::value && std::is_move_assignable::value ? std::is_trivially_move_assignable::value || std::is_nothrow_move_assignable::value : std::is_trivially_copy_assignable::value || std::is_nothrow_copy_assignable::value) && MOODYCAMEL_NOEXCEPT_CTOR(type, valueType, expr)) +#elif defined(_MSC_VER) && defined(_NOEXCEPT) && _MSC_VER < 1900 +#define MOODYCAMEL_NOEXCEPT _NOEXCEPT +#define MOODYCAMEL_NOEXCEPT_CTOR(type, valueType, expr) (std::is_rvalue_reference::value && std::is_move_constructible::value ? std::is_trivially_move_constructible::value || std::is_nothrow_move_constructible::value : std::is_trivially_copy_constructible::value || std::is_nothrow_copy_constructible::value) +#define MOODYCAMEL_NOEXCEPT_ASSIGN(type, valueType, expr) ((std::is_rvalue_reference::value && std::is_move_assignable::value ? std::is_trivially_move_assignable::value || std::is_nothrow_move_assignable::value : std::is_trivially_copy_assignable::value || std::is_nothrow_copy_assignable::value) && MOODYCAMEL_NOEXCEPT_CTOR(type, valueType, expr)) +#else +#define MOODYCAMEL_NOEXCEPT noexcept +#define MOODYCAMEL_NOEXCEPT_CTOR(type, valueType, expr) noexcept(expr) +#define MOODYCAMEL_NOEXCEPT_ASSIGN(type, valueType, expr) noexcept(expr) +#endif +#endif + +#ifndef MOODYCAMEL_CPP11_THREAD_LOCAL_SUPPORTED +#ifdef MCDBGQ_USE_RELACY +#define MOODYCAMEL_CPP11_THREAD_LOCAL_SUPPORTED +#else +// VS2013 doesn't support `thread_local`, and MinGW-w64 w/ POSIX threading has a crippling bug: http://sourceforge.net/p/mingw-w64/bugs/445 +// g++ <=4.7 doesn't support thread_local either. +// Finally, iOS/ARM doesn't have support for it either, and g++/ARM allows it to compile but it's unconfirmed to actually work +#if (!defined(_MSC_VER) || _MSC_VER >= 1900) && (!defined(__MINGW32__) && !defined(__MINGW64__) || !defined(__WINPTHREADS_VERSION)) && (!defined(__GNUC__) || __GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 8)) && (!defined(__APPLE__) || !TARGET_OS_IPHONE) && !defined(__arm__) && !defined(_M_ARM) && !defined(__aarch64__) +// Assume `thread_local` is fully supported in all other C++11 compilers/platforms +//#define MOODYCAMEL_CPP11_THREAD_LOCAL_SUPPORTED // always disabled for now since several users report having problems with it on +#endif +#endif +#endif + +// VS2012 doesn't support deleted functions. +// In this case, we declare the function normally but don't define it. A link error will be generated if the function is called. +#ifndef MOODYCAMEL_DELETE_FUNCTION +#if defined(_MSC_VER) && _MSC_VER < 1800 +#define MOODYCAMEL_DELETE_FUNCTION +#else +#define MOODYCAMEL_DELETE_FUNCTION = delete +#endif +#endif + +// Compiler-specific likely/unlikely hints +namespace moodycamel { namespace details { +#if defined(__GNUC__) + static inline bool (likely)(bool x) { return __builtin_expect((x), true); } + static inline bool (unlikely)(bool x) { return __builtin_expect((x), false); } +#else + static inline bool (likely)(bool x) { return x; } + static inline bool (unlikely)(bool x) { return x; } +#endif +} } + +#ifdef MOODYCAMEL_QUEUE_INTERNAL_DEBUG +#include "internal/concurrentqueue_internal_debug.h" +#endif + +namespace moodycamel { +namespace details { + template + struct const_numeric_max { + static_assert(std::is_integral::value, "const_numeric_max can only be used with integers"); + static const T value = std::numeric_limits::is_signed + ? (static_cast(1) << (sizeof(T) * CHAR_BIT - 1)) - static_cast(1) + : static_cast(-1); + }; + +#if defined(__GLIBCXX__) + typedef ::max_align_t std_max_align_t; // libstdc++ forgot to add it to std:: for a while +#else + typedef std::max_align_t std_max_align_t; // Others (e.g. MSVC) insist it can *only* be accessed via std:: +#endif + + // Some platforms have incorrectly set max_align_t to a type with <8 bytes alignment even while supporting + // 8-byte aligned scalar values (*cough* 32-bit iOS). Work around this with our own union. See issue #64. + typedef union { + std_max_align_t x; + long long y; + void* z; + } max_align_t; +} + +// Default traits for the ConcurrentQueue. To change some of the +// traits without re-implementing all of them, inherit from this +// struct and shadow the declarations you wish to be different; +// since the traits are used as a template type parameter, the +// shadowed declarations will be used where defined, and the defaults +// otherwise. +struct ConcurrentQueueDefaultTraits +{ + // General-purpose size type. std::size_t is strongly recommended. + typedef std::size_t size_t; + + // The type used for the enqueue and dequeue indices. Must be at least as + // large as size_t. Should be significantly larger than the number of elements + // you expect to hold at once, especially if you have a high turnover rate; + // for example, on 32-bit x86, if you expect to have over a hundred million + // elements or pump several million elements through your queue in a very + // short space of time, using a 32-bit type *may* trigger a race condition. + // A 64-bit int type is recommended in that case, and in practice will + // prevent a race condition no matter the usage of the queue. Note that + // whether the queue is lock-free with a 64-int type depends on the whether + // std::atomic is lock-free, which is platform-specific. + typedef std::size_t index_t; + + // Internally, all elements are enqueued and dequeued from multi-element + // blocks; this is the smallest controllable unit. If you expect few elements + // but many producers, a smaller block size should be favoured. For few producers + // and/or many elements, a larger block size is preferred. A sane default + // is provided. Must be a power of 2. + static const size_t BLOCK_SIZE = 32; + + // For explicit producers (i.e. when using a producer token), the block is + // checked for being empty by iterating through a list of flags, one per element. + // For large block sizes, this is too inefficient, and switching to an atomic + // counter-based approach is faster. The switch is made for block sizes strictly + // larger than this threshold. + static const size_t EXPLICIT_BLOCK_EMPTY_COUNTER_THRESHOLD = 32; + + // How many full blocks can be expected for a single explicit producer? This should + // reflect that number's maximum for optimal performance. Must be a power of 2. + static const size_t EXPLICIT_INITIAL_INDEX_SIZE = 32; + + // How many full blocks can be expected for a single implicit producer? This should + // reflect that number's maximum for optimal performance. Must be a power of 2. + static const size_t IMPLICIT_INITIAL_INDEX_SIZE = 32; + + // The initial size of the hash table mapping thread IDs to implicit producers. + // Note that the hash is resized every time it becomes half full. + // Must be a power of two, and either 0 or at least 1. If 0, implicit production + // (using the enqueue methods without an explicit producer token) is disabled. + static const size_t INITIAL_IMPLICIT_PRODUCER_HASH_SIZE = 32; + + // Controls the number of items that an explicit consumer (i.e. one with a token) + // must consume before it causes all consumers to rotate and move on to the next + // internal queue. + static const std::uint32_t EXPLICIT_CONSUMER_CONSUMPTION_QUOTA_BEFORE_ROTATE = 256; + + // The maximum number of elements (inclusive) that can be enqueued to a sub-queue. + // Enqueue operations that would cause this limit to be surpassed will fail. Note + // that this limit is enforced at the block level (for performance reasons), i.e. + // it's rounded up to the nearest block size. + static const size_t MAX_SUBQUEUE_SIZE = details::const_numeric_max::value; + + +#ifndef MCDBGQ_USE_RELACY + // Memory allocation can be customized if needed. + // malloc should return nullptr on failure, and handle alignment like std::malloc. +#if defined(malloc) || defined(free) + // Gah, this is 2015, stop defining macros that break standard code already! + // Work around malloc/free being special macros: + static inline void* WORKAROUND_malloc(size_t size) { return malloc(size); } + static inline void WORKAROUND_free(void* ptr) { return free(ptr); } + static inline void* (malloc)(size_t size) { return WORKAROUND_malloc(size); } + static inline void (free)(void* ptr) { return WORKAROUND_free(ptr); } +#else + static inline void* malloc(size_t size) { return std::malloc(size); } + static inline void free(void* ptr) { return std::free(ptr); } +#endif +#else + // Debug versions when running under the Relacy race detector (ignore + // these in user code) + static inline void* malloc(size_t size) { return rl::rl_malloc(size, $); } + static inline void free(void* ptr) { return rl::rl_free(ptr, $); } +#endif +}; + + +// When producing or consuming many elements, the most efficient way is to: +// 1) Use one of the bulk-operation methods of the queue with a token +// 2) Failing that, use the bulk-operation methods without a token +// 3) Failing that, create a token and use that with the single-item methods +// 4) Failing that, use the single-parameter methods of the queue +// Having said that, don't create tokens willy-nilly -- ideally there should be +// a maximum of one token per thread (of each kind). +struct ProducerToken; +struct ConsumerToken; + +template class ConcurrentQueue; +template class BlockingConcurrentQueue; +class ConcurrentQueueTests; + + +namespace details +{ + struct ConcurrentQueueProducerTypelessBase + { + ConcurrentQueueProducerTypelessBase* next; + std::atomic inactive; + ProducerToken* token; + + ConcurrentQueueProducerTypelessBase() + : next(nullptr), inactive(false), token(nullptr) + { + } + }; + + template struct _hash_32_or_64 { + static inline std::uint32_t hash(std::uint32_t h) + { + // MurmurHash3 finalizer -- see https://code.google.com/p/smhasher/source/browse/trunk/MurmurHash3.cpp + // Since the thread ID is already unique, all we really want to do is propagate that + // uniqueness evenly across all the bits, so that we can use a subset of the bits while + // reducing collisions significantly + h ^= h >> 16; + h *= 0x85ebca6b; + h ^= h >> 13; + h *= 0xc2b2ae35; + return h ^ (h >> 16); + } + }; + template<> struct _hash_32_or_64<1> { + static inline std::uint64_t hash(std::uint64_t h) + { + h ^= h >> 33; + h *= 0xff51afd7ed558ccd; + h ^= h >> 33; + h *= 0xc4ceb9fe1a85ec53; + return h ^ (h >> 33); + } + }; + template struct hash_32_or_64 : public _hash_32_or_64<(size > 4)> { }; + + static inline size_t hash_thread_id(thread_id_t id) + { + static_assert(sizeof(thread_id_t) <= 8, "Expected a platform where thread IDs are at most 64-bit values"); + return static_cast(hash_32_or_64::thread_id_hash_t)>::hash( + thread_id_converter::prehash(id))); + } + + template + static inline bool circular_less_than(T a, T b) + { +#ifdef _MSC_VER +#pragma warning(push) +#pragma warning(disable: 4554) +#endif + static_assert(std::is_integral::value && !std::numeric_limits::is_signed, "circular_less_than is intended to be used only with unsigned integer types"); + return static_cast(a - b) > static_cast(static_cast(1) << static_cast(sizeof(T) * CHAR_BIT - 1)); +#ifdef _MSC_VER +#pragma warning(pop) +#endif + } + + template + static inline char* align_for(char* ptr) + { + const std::size_t alignment = std::alignment_of::value; + return ptr + (alignment - (reinterpret_cast(ptr) % alignment)) % alignment; + } + + template + static inline T ceil_to_pow_2(T x) + { + static_assert(std::is_integral::value && !std::numeric_limits::is_signed, "ceil_to_pow_2 is intended to be used only with unsigned integer types"); + + // Adapted from http://graphics.stanford.edu/~seander/bithacks.html#RoundUpPowerOf2 + --x; + x |= x >> 1; + x |= x >> 2; + x |= x >> 4; + for (std::size_t i = 1; i < sizeof(T); i <<= 1) { + x |= x >> (i << 3); + } + ++x; + return x; + } + + template + static inline void swap_relaxed(std::atomic& left, std::atomic& right) + { + T temp = std::move(left.load(std::memory_order_relaxed)); + left.store(std::move(right.load(std::memory_order_relaxed)), std::memory_order_relaxed); + right.store(std::move(temp), std::memory_order_relaxed); + } + + template + static inline T const& nomove(T const& x) + { + return x; + } + + template + struct nomove_if + { + template + static inline T const& eval(T const& x) + { + return x; + } + }; + + template<> + struct nomove_if + { + template + static inline auto eval(U&& x) + -> decltype(std::forward(x)) + { + return std::forward(x); + } + }; + + template + static inline auto deref_noexcept(It& it) MOODYCAMEL_NOEXCEPT -> decltype(*it) + { + return *it; + } + +#if defined(__clang__) || !defined(__GNUC__) || __GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 8) + template struct is_trivially_destructible : std::is_trivially_destructible { }; +#else + template struct is_trivially_destructible : std::has_trivial_destructor { }; +#endif + +#ifdef MOODYCAMEL_CPP11_THREAD_LOCAL_SUPPORTED +#ifdef MCDBGQ_USE_RELACY + typedef RelacyThreadExitListener ThreadExitListener; + typedef RelacyThreadExitNotifier ThreadExitNotifier; +#else + struct ThreadExitListener + { + typedef void (*callback_t)(void*); + callback_t callback; + void* userData; + + ThreadExitListener* next; // reserved for use by the ThreadExitNotifier + }; + + + class ThreadExitNotifier + { + public: + static void subscribe(ThreadExitListener* listener) + { + auto& tlsInst = instance(); + listener->next = tlsInst.tail; + tlsInst.tail = listener; + } + + static void unsubscribe(ThreadExitListener* listener) + { + auto& tlsInst = instance(); + ThreadExitListener** prev = &tlsInst.tail; + for (auto ptr = tlsInst.tail; ptr != nullptr; ptr = ptr->next) { + if (ptr == listener) { + *prev = ptr->next; + break; + } + prev = &ptr->next; + } + } + + private: + ThreadExitNotifier() : tail(nullptr) { } + ThreadExitNotifier(ThreadExitNotifier const&) MOODYCAMEL_DELETE_FUNCTION; + ThreadExitNotifier& operator=(ThreadExitNotifier const&) MOODYCAMEL_DELETE_FUNCTION; + + ~ThreadExitNotifier() + { + // This thread is about to exit, let everyone know! + assert(this == &instance() && "If this assert fails, you likely have a buggy compiler! Change the preprocessor conditions such that MOODYCAMEL_CPP11_THREAD_LOCAL_SUPPORTED is no longer defined."); + for (auto ptr = tail; ptr != nullptr; ptr = ptr->next) { + ptr->callback(ptr->userData); + } + } + + // Thread-local + static inline ThreadExitNotifier& instance() + { + static thread_local ThreadExitNotifier notifier; + return notifier; + } + + private: + ThreadExitListener* tail; + }; +#endif +#endif + + template struct static_is_lock_free_num { enum { value = 0 }; }; + template<> struct static_is_lock_free_num { enum { value = ATOMIC_CHAR_LOCK_FREE }; }; + template<> struct static_is_lock_free_num { enum { value = ATOMIC_SHORT_LOCK_FREE }; }; + template<> struct static_is_lock_free_num { enum { value = ATOMIC_INT_LOCK_FREE }; }; + template<> struct static_is_lock_free_num { enum { value = ATOMIC_LONG_LOCK_FREE }; }; + template<> struct static_is_lock_free_num { enum { value = ATOMIC_LLONG_LOCK_FREE }; }; + template struct static_is_lock_free : static_is_lock_free_num::type> { }; + template<> struct static_is_lock_free { enum { value = ATOMIC_BOOL_LOCK_FREE }; }; + template struct static_is_lock_free { enum { value = ATOMIC_POINTER_LOCK_FREE }; }; +} + + +struct ProducerToken +{ + template + explicit ProducerToken(ConcurrentQueue& queue); + + template + explicit ProducerToken(BlockingConcurrentQueue& queue); + + ProducerToken(ProducerToken&& other) MOODYCAMEL_NOEXCEPT + : producer(other.producer) + { + other.producer = nullptr; + if (producer != nullptr) { + producer->token = this; + } + } + + inline ProducerToken& operator=(ProducerToken&& other) MOODYCAMEL_NOEXCEPT + { + swap(other); + return *this; + } + + void swap(ProducerToken& other) MOODYCAMEL_NOEXCEPT + { + std::swap(producer, other.producer); + if (producer != nullptr) { + producer->token = this; + } + if (other.producer != nullptr) { + other.producer->token = &other; + } + } + + // A token is always valid unless: + // 1) Memory allocation failed during construction + // 2) It was moved via the move constructor + // (Note: assignment does a swap, leaving both potentially valid) + // 3) The associated queue was destroyed + // Note that if valid() returns true, that only indicates + // that the token is valid for use with a specific queue, + // but not which one; that's up to the user to track. + inline bool valid() const { return producer != nullptr; } + + ~ProducerToken() + { + if (producer != nullptr) { + producer->token = nullptr; + producer->inactive.store(true, std::memory_order_release); + } + } + + // Disable copying and assignment + ProducerToken(ProducerToken const&) MOODYCAMEL_DELETE_FUNCTION; + ProducerToken& operator=(ProducerToken const&) MOODYCAMEL_DELETE_FUNCTION; + +private: + template friend class ConcurrentQueue; + friend class ConcurrentQueueTests; + +protected: + details::ConcurrentQueueProducerTypelessBase* producer; +}; + + +struct ConsumerToken +{ + template + explicit ConsumerToken(ConcurrentQueue& q); + + template + explicit ConsumerToken(BlockingConcurrentQueue& q); + + ConsumerToken(ConsumerToken&& other) MOODYCAMEL_NOEXCEPT + : initialOffset(other.initialOffset), lastKnownGlobalOffset(other.lastKnownGlobalOffset), itemsConsumedFromCurrent(other.itemsConsumedFromCurrent), currentProducer(other.currentProducer), desiredProducer(other.desiredProducer) + { + } + + inline ConsumerToken& operator=(ConsumerToken&& other) MOODYCAMEL_NOEXCEPT + { + swap(other); + return *this; + } + + void swap(ConsumerToken& other) MOODYCAMEL_NOEXCEPT + { + std::swap(initialOffset, other.initialOffset); + std::swap(lastKnownGlobalOffset, other.lastKnownGlobalOffset); + std::swap(itemsConsumedFromCurrent, other.itemsConsumedFromCurrent); + std::swap(currentProducer, other.currentProducer); + std::swap(desiredProducer, other.desiredProducer); + } + + // Disable copying and assignment + ConsumerToken(ConsumerToken const&) MOODYCAMEL_DELETE_FUNCTION; + ConsumerToken& operator=(ConsumerToken const&) MOODYCAMEL_DELETE_FUNCTION; + +private: + template friend class ConcurrentQueue; + friend class ConcurrentQueueTests; + +private: // but shared with ConcurrentQueue + std::uint32_t initialOffset; + std::uint32_t lastKnownGlobalOffset; + std::uint32_t itemsConsumedFromCurrent; + details::ConcurrentQueueProducerTypelessBase* currentProducer; + details::ConcurrentQueueProducerTypelessBase* desiredProducer; +}; + +// Need to forward-declare this swap because it's in a namespace. +// See http://stackoverflow.com/questions/4492062/why-does-a-c-friend-class-need-a-forward-declaration-only-in-other-namespaces +template +inline void swap(typename ConcurrentQueue::ImplicitProducerKVP& a, typename ConcurrentQueue::ImplicitProducerKVP& b) MOODYCAMEL_NOEXCEPT; + + +template +class ConcurrentQueue +{ +public: + typedef ::moodycamel::ProducerToken producer_token_t; + typedef ::moodycamel::ConsumerToken consumer_token_t; + + typedef typename Traits::index_t index_t; + typedef typename Traits::size_t size_t; + + static const size_t BLOCK_SIZE = static_cast(Traits::BLOCK_SIZE); + static const size_t EXPLICIT_BLOCK_EMPTY_COUNTER_THRESHOLD = static_cast(Traits::EXPLICIT_BLOCK_EMPTY_COUNTER_THRESHOLD); + static const size_t EXPLICIT_INITIAL_INDEX_SIZE = static_cast(Traits::EXPLICIT_INITIAL_INDEX_SIZE); + static const size_t IMPLICIT_INITIAL_INDEX_SIZE = static_cast(Traits::IMPLICIT_INITIAL_INDEX_SIZE); + static const size_t INITIAL_IMPLICIT_PRODUCER_HASH_SIZE = static_cast(Traits::INITIAL_IMPLICIT_PRODUCER_HASH_SIZE); + static const std::uint32_t EXPLICIT_CONSUMER_CONSUMPTION_QUOTA_BEFORE_ROTATE = static_cast(Traits::EXPLICIT_CONSUMER_CONSUMPTION_QUOTA_BEFORE_ROTATE); +#ifdef _MSC_VER +#pragma warning(push) +#pragma warning(disable: 4307) // + integral constant overflow (that's what the ternary expression is for!) +#pragma warning(disable: 4309) // static_cast: Truncation of constant value +#endif + static const size_t MAX_SUBQUEUE_SIZE = (details::const_numeric_max::value - static_cast(Traits::MAX_SUBQUEUE_SIZE) < BLOCK_SIZE) ? details::const_numeric_max::value : ((static_cast(Traits::MAX_SUBQUEUE_SIZE) + (BLOCK_SIZE - 1)) / BLOCK_SIZE * BLOCK_SIZE); +#ifdef _MSC_VER +#pragma warning(pop) +#endif + + static_assert(!std::numeric_limits::is_signed && std::is_integral::value, "Traits::size_t must be an unsigned integral type"); + static_assert(!std::numeric_limits::is_signed && std::is_integral::value, "Traits::index_t must be an unsigned integral type"); + static_assert(sizeof(index_t) >= sizeof(size_t), "Traits::index_t must be at least as wide as Traits::size_t"); + static_assert((BLOCK_SIZE > 1) && !(BLOCK_SIZE & (BLOCK_SIZE - 1)), "Traits::BLOCK_SIZE must be a power of 2 (and at least 2)"); + static_assert((EXPLICIT_BLOCK_EMPTY_COUNTER_THRESHOLD > 1) && !(EXPLICIT_BLOCK_EMPTY_COUNTER_THRESHOLD & (EXPLICIT_BLOCK_EMPTY_COUNTER_THRESHOLD - 1)), "Traits::EXPLICIT_BLOCK_EMPTY_COUNTER_THRESHOLD must be a power of 2 (and greater than 1)"); + static_assert((EXPLICIT_INITIAL_INDEX_SIZE > 1) && !(EXPLICIT_INITIAL_INDEX_SIZE & (EXPLICIT_INITIAL_INDEX_SIZE - 1)), "Traits::EXPLICIT_INITIAL_INDEX_SIZE must be a power of 2 (and greater than 1)"); + static_assert((IMPLICIT_INITIAL_INDEX_SIZE > 1) && !(IMPLICIT_INITIAL_INDEX_SIZE & (IMPLICIT_INITIAL_INDEX_SIZE - 1)), "Traits::IMPLICIT_INITIAL_INDEX_SIZE must be a power of 2 (and greater than 1)"); + static_assert((INITIAL_IMPLICIT_PRODUCER_HASH_SIZE == 0) || !(INITIAL_IMPLICIT_PRODUCER_HASH_SIZE & (INITIAL_IMPLICIT_PRODUCER_HASH_SIZE - 1)), "Traits::INITIAL_IMPLICIT_PRODUCER_HASH_SIZE must be a power of 2"); + static_assert(INITIAL_IMPLICIT_PRODUCER_HASH_SIZE == 0 || INITIAL_IMPLICIT_PRODUCER_HASH_SIZE >= 1, "Traits::INITIAL_IMPLICIT_PRODUCER_HASH_SIZE must be at least 1 (or 0 to disable implicit enqueueing)"); + +public: + // Creates a queue with at least `capacity` element slots; note that the + // actual number of elements that can be inserted without additional memory + // allocation depends on the number of producers and the block size (e.g. if + // the block size is equal to `capacity`, only a single block will be allocated + // up-front, which means only a single producer will be able to enqueue elements + // without an extra allocation -- blocks aren't shared between producers). + // This method is not thread safe -- it is up to the user to ensure that the + // queue is fully constructed before it starts being used by other threads (this + // includes making the memory effects of construction visible, possibly with a + // memory barrier). + explicit ConcurrentQueue(size_t capacity = 6 * BLOCK_SIZE) + : producerListTail(nullptr), + producerCount(0), + initialBlockPoolIndex(0), + nextExplicitConsumerId(0), + globalExplicitConsumerOffset(0) + { + implicitProducerHashResizeInProgress.clear(std::memory_order_relaxed); + populate_initial_implicit_producer_hash(); + populate_initial_block_list(capacity / BLOCK_SIZE + ((capacity & (BLOCK_SIZE - 1)) == 0 ? 0 : 1)); + +#ifdef MOODYCAMEL_QUEUE_INTERNAL_DEBUG + // Track all the producers using a fully-resolved typed list for + // each kind; this makes it possible to debug them starting from + // the root queue object (otherwise wacky casts are needed that + // don't compile in the debugger's expression evaluator). + explicitProducers.store(nullptr, std::memory_order_relaxed); + implicitProducers.store(nullptr, std::memory_order_relaxed); +#endif + } + + // Computes the correct amount of pre-allocated blocks for you based + // on the minimum number of elements you want available at any given + // time, and the maximum concurrent number of each type of producer. + ConcurrentQueue(size_t minCapacity, size_t maxExplicitProducers, size_t maxImplicitProducers) + : producerListTail(nullptr), + producerCount(0), + initialBlockPoolIndex(0), + nextExplicitConsumerId(0), + globalExplicitConsumerOffset(0) + { + implicitProducerHashResizeInProgress.clear(std::memory_order_relaxed); + populate_initial_implicit_producer_hash(); + size_t blocks = (((minCapacity + BLOCK_SIZE - 1) / BLOCK_SIZE) - 1) * (maxExplicitProducers + 1) + 2 * (maxExplicitProducers + maxImplicitProducers); + populate_initial_block_list(blocks); + +#ifdef MOODYCAMEL_QUEUE_INTERNAL_DEBUG + explicitProducers.store(nullptr, std::memory_order_relaxed); + implicitProducers.store(nullptr, std::memory_order_relaxed); +#endif + } + + // Note: The queue should not be accessed concurrently while it's + // being deleted. It's up to the user to synchronize this. + // This method is not thread safe. + ~ConcurrentQueue() + { + // Destroy producers + auto ptr = producerListTail.load(std::memory_order_relaxed); + while (ptr != nullptr) { + auto next = ptr->next_prod(); + if (ptr->token != nullptr) { + ptr->token->producer = nullptr; + } + destroy(ptr); + ptr = next; + } + + // Destroy implicit producer hash tables + if (INITIAL_IMPLICIT_PRODUCER_HASH_SIZE != 0) { + auto hash = implicitProducerHash.load(std::memory_order_relaxed); + while (hash != nullptr) { + auto prev = hash->prev; + if (prev != nullptr) { // The last hash is part of this object and was not allocated dynamically + for (size_t i = 0; i != hash->capacity; ++i) { + hash->entries[i].~ImplicitProducerKVP(); + } + hash->~ImplicitProducerHash(); + (Traits::free)(hash); + } + hash = prev; + } + } + + // Destroy global free list + auto block = freeList.head_unsafe(); + while (block != nullptr) { + auto next = block->freeListNext.load(std::memory_order_relaxed); + if (block->dynamicallyAllocated) { + destroy(block); + } + block = next; + } + + // Destroy initial free list + destroy_array(initialBlockPool, initialBlockPoolSize); + } + + // Disable copying and copy assignment + ConcurrentQueue(ConcurrentQueue const&) MOODYCAMEL_DELETE_FUNCTION; + ConcurrentQueue& operator=(ConcurrentQueue const&) MOODYCAMEL_DELETE_FUNCTION; + + // Moving is supported, but note that it is *not* a thread-safe operation. + // Nobody can use the queue while it's being moved, and the memory effects + // of that move must be propagated to other threads before they can use it. + // Note: When a queue is moved, its tokens are still valid but can only be + // used with the destination queue (i.e. semantically they are moved along + // with the queue itself). + ConcurrentQueue(ConcurrentQueue&& other) MOODYCAMEL_NOEXCEPT + : producerListTail(other.producerListTail.load(std::memory_order_relaxed)), + producerCount(other.producerCount.load(std::memory_order_relaxed)), + initialBlockPoolIndex(other.initialBlockPoolIndex.load(std::memory_order_relaxed)), + initialBlockPool(other.initialBlockPool), + initialBlockPoolSize(other.initialBlockPoolSize), + freeList(std::move(other.freeList)), + nextExplicitConsumerId(other.nextExplicitConsumerId.load(std::memory_order_relaxed)), + globalExplicitConsumerOffset(other.globalExplicitConsumerOffset.load(std::memory_order_relaxed)) + { + // Move the other one into this, and leave the other one as an empty queue + implicitProducerHashResizeInProgress.clear(std::memory_order_relaxed); + populate_initial_implicit_producer_hash(); + swap_implicit_producer_hashes(other); + + other.producerListTail.store(nullptr, std::memory_order_relaxed); + other.producerCount.store(0, std::memory_order_relaxed); + other.nextExplicitConsumerId.store(0, std::memory_order_relaxed); + other.globalExplicitConsumerOffset.store(0, std::memory_order_relaxed); + +#ifdef MOODYCAMEL_QUEUE_INTERNAL_DEBUG + explicitProducers.store(other.explicitProducers.load(std::memory_order_relaxed), std::memory_order_relaxed); + other.explicitProducers.store(nullptr, std::memory_order_relaxed); + implicitProducers.store(other.implicitProducers.load(std::memory_order_relaxed), std::memory_order_relaxed); + other.implicitProducers.store(nullptr, std::memory_order_relaxed); +#endif + + other.initialBlockPoolIndex.store(0, std::memory_order_relaxed); + other.initialBlockPoolSize = 0; + other.initialBlockPool = nullptr; + + reown_producers(); + } + + inline ConcurrentQueue& operator=(ConcurrentQueue&& other) MOODYCAMEL_NOEXCEPT + { + return swap_internal(other); + } + + // Swaps this queue's state with the other's. Not thread-safe. + // Swapping two queues does not invalidate their tokens, however + // the tokens that were created for one queue must be used with + // only the swapped queue (i.e. the tokens are tied to the + // queue's movable state, not the object itself). + inline void swap(ConcurrentQueue& other) MOODYCAMEL_NOEXCEPT + { + swap_internal(other); + } + +private: + ConcurrentQueue& swap_internal(ConcurrentQueue& other) + { + if (this == &other) { + return *this; + } + + details::swap_relaxed(producerListTail, other.producerListTail); + details::swap_relaxed(producerCount, other.producerCount); + details::swap_relaxed(initialBlockPoolIndex, other.initialBlockPoolIndex); + std::swap(initialBlockPool, other.initialBlockPool); + std::swap(initialBlockPoolSize, other.initialBlockPoolSize); + freeList.swap(other.freeList); + details::swap_relaxed(nextExplicitConsumerId, other.nextExplicitConsumerId); + details::swap_relaxed(globalExplicitConsumerOffset, other.globalExplicitConsumerOffset); + + swap_implicit_producer_hashes(other); + + reown_producers(); + other.reown_producers(); + +#ifdef MOODYCAMEL_QUEUE_INTERNAL_DEBUG + details::swap_relaxed(explicitProducers, other.explicitProducers); + details::swap_relaxed(implicitProducers, other.implicitProducers); +#endif + + return *this; + } + +public: + // Enqueues a single item (by copying it). + // Allocates memory if required. Only fails if memory allocation fails (or implicit + // production is disabled because Traits::INITIAL_IMPLICIT_PRODUCER_HASH_SIZE is 0, + // or Traits::MAX_SUBQUEUE_SIZE has been defined and would be surpassed). + // Thread-safe. + inline bool enqueue(T const& item) + { + if (INITIAL_IMPLICIT_PRODUCER_HASH_SIZE == 0) return false; + return inner_enqueue(item); + } + + // Enqueues a single item (by moving it, if possible). + // Allocates memory if required. Only fails if memory allocation fails (or implicit + // production is disabled because Traits::INITIAL_IMPLICIT_PRODUCER_HASH_SIZE is 0, + // or Traits::MAX_SUBQUEUE_SIZE has been defined and would be surpassed). + // Thread-safe. + inline bool enqueue(T&& item) + { + if (INITIAL_IMPLICIT_PRODUCER_HASH_SIZE == 0) return false; + return inner_enqueue(std::move(item)); + } + + // Enqueues a single item (by copying it) using an explicit producer token. + // Allocates memory if required. Only fails if memory allocation fails (or + // Traits::MAX_SUBQUEUE_SIZE has been defined and would be surpassed). + // Thread-safe. + inline bool enqueue(producer_token_t const& token, T const& item) + { + return inner_enqueue(token, item); + } + + // Enqueues a single item (by moving it, if possible) using an explicit producer token. + // Allocates memory if required. Only fails if memory allocation fails (or + // Traits::MAX_SUBQUEUE_SIZE has been defined and would be surpassed). + // Thread-safe. + inline bool enqueue(producer_token_t const& token, T&& item) + { + return inner_enqueue(token, std::move(item)); + } + + // Enqueues several items. + // Allocates memory if required. Only fails if memory allocation fails (or + // implicit production is disabled because Traits::INITIAL_IMPLICIT_PRODUCER_HASH_SIZE + // is 0, or Traits::MAX_SUBQUEUE_SIZE has been defined and would be surpassed). + // Note: Use std::make_move_iterator if the elements should be moved instead of copied. + // Thread-safe. + template + bool enqueue_bulk(It itemFirst, size_t count) + { + if (INITIAL_IMPLICIT_PRODUCER_HASH_SIZE == 0) return false; + return inner_enqueue_bulk(itemFirst, count); + } + + // Enqueues several items using an explicit producer token. + // Allocates memory if required. Only fails if memory allocation fails + // (or Traits::MAX_SUBQUEUE_SIZE has been defined and would be surpassed). + // Note: Use std::make_move_iterator if the elements should be moved + // instead of copied. + // Thread-safe. + template + bool enqueue_bulk(producer_token_t const& token, It itemFirst, size_t count) + { + return inner_enqueue_bulk(token, itemFirst, count); + } + + // Enqueues a single item (by copying it). + // Does not allocate memory. Fails if not enough room to enqueue (or implicit + // production is disabled because Traits::INITIAL_IMPLICIT_PRODUCER_HASH_SIZE + // is 0). + // Thread-safe. + inline bool try_enqueue(T const& item) + { + if (INITIAL_IMPLICIT_PRODUCER_HASH_SIZE == 0) return false; + return inner_enqueue(item); + } + + // Enqueues a single item (by moving it, if possible). + // Does not allocate memory (except for one-time implicit producer). + // Fails if not enough room to enqueue (or implicit production is + // disabled because Traits::INITIAL_IMPLICIT_PRODUCER_HASH_SIZE is 0). + // Thread-safe. + inline bool try_enqueue(T&& item) + { + if (INITIAL_IMPLICIT_PRODUCER_HASH_SIZE == 0) return false; + return inner_enqueue(std::move(item)); + } + + // Enqueues a single item (by copying it) using an explicit producer token. + // Does not allocate memory. Fails if not enough room to enqueue. + // Thread-safe. + inline bool try_enqueue(producer_token_t const& token, T const& item) + { + return inner_enqueue(token, item); + } + + // Enqueues a single item (by moving it, if possible) using an explicit producer token. + // Does not allocate memory. Fails if not enough room to enqueue. + // Thread-safe. + inline bool try_enqueue(producer_token_t const& token, T&& item) + { + return inner_enqueue(token, std::move(item)); + } + + // Enqueues several items. + // Does not allocate memory (except for one-time implicit producer). + // Fails if not enough room to enqueue (or implicit production is + // disabled because Traits::INITIAL_IMPLICIT_PRODUCER_HASH_SIZE is 0). + // Note: Use std::make_move_iterator if the elements should be moved + // instead of copied. + // Thread-safe. + template + bool try_enqueue_bulk(It itemFirst, size_t count) + { + if (INITIAL_IMPLICIT_PRODUCER_HASH_SIZE == 0) return false; + return inner_enqueue_bulk(itemFirst, count); + } + + // Enqueues several items using an explicit producer token. + // Does not allocate memory. Fails if not enough room to enqueue. + // Note: Use std::make_move_iterator if the elements should be moved + // instead of copied. + // Thread-safe. + template + bool try_enqueue_bulk(producer_token_t const& token, It itemFirst, size_t count) + { + return inner_enqueue_bulk(token, itemFirst, count); + } + + + + // Attempts to dequeue from the queue. + // Returns false if all producer streams appeared empty at the time they + // were checked (so, the queue is likely but not guaranteed to be empty). + // Never allocates. Thread-safe. + template + bool try_dequeue(U& item) + { + // Instead of simply trying each producer in turn (which could cause needless contention on the first + // producer), we score them heuristically. + size_t nonEmptyCount = 0; + ProducerBase* best = nullptr; + size_t bestSize = 0; + for (auto ptr = producerListTail.load(std::memory_order_acquire); nonEmptyCount < 3 && ptr != nullptr; ptr = ptr->next_prod()) { + auto size = ptr->size_approx(); + if (size > 0) { + if (size > bestSize) { + bestSize = size; + best = ptr; + } + ++nonEmptyCount; + } + } + + // If there was at least one non-empty queue but it appears empty at the time + // we try to dequeue from it, we need to make sure every queue's been tried + if (nonEmptyCount > 0) { + if ((details::likely)(best->dequeue(item))) { + return true; + } + for (auto ptr = producerListTail.load(std::memory_order_acquire); ptr != nullptr; ptr = ptr->next_prod()) { + if (ptr != best && ptr->dequeue(item)) { + return true; + } + } + } + return false; + } + + // Attempts to dequeue from the queue. + // Returns false if all producer streams appeared empty at the time they + // were checked (so, the queue is likely but not guaranteed to be empty). + // This differs from the try_dequeue(item) method in that this one does + // not attempt to reduce contention by interleaving the order that producer + // streams are dequeued from. So, using this method can reduce overall throughput + // under contention, but will give more predictable results in single-threaded + // consumer scenarios. This is mostly only useful for internal unit tests. + // Never allocates. Thread-safe. + template + bool try_dequeue_non_interleaved(U& item) + { + for (auto ptr = producerListTail.load(std::memory_order_acquire); ptr != nullptr; ptr = ptr->next_prod()) { + if (ptr->dequeue(item)) { + return true; + } + } + return false; + } + + // Attempts to dequeue from the queue using an explicit consumer token. + // Returns false if all producer streams appeared empty at the time they + // were checked (so, the queue is likely but not guaranteed to be empty). + // Never allocates. Thread-safe. + template + bool try_dequeue(consumer_token_t& token, U& item) + { + // The idea is roughly as follows: + // Every 256 items from one producer, make everyone rotate (increase the global offset) -> this means the highest efficiency consumer dictates the rotation speed of everyone else, more or less + // If you see that the global offset has changed, you must reset your consumption counter and move to your designated place + // If there's no items where you're supposed to be, keep moving until you find a producer with some items + // If the global offset has not changed but you've run out of items to consume, move over from your current position until you find an producer with something in it + + if (token.desiredProducer == nullptr || token.lastKnownGlobalOffset != globalExplicitConsumerOffset.load(std::memory_order_relaxed)) { + if (!update_current_producer_after_rotation(token)) { + return false; + } + } + + // If there was at least one non-empty queue but it appears empty at the time + // we try to dequeue from it, we need to make sure every queue's been tried + if (static_cast(token.currentProducer)->dequeue(item)) { + if (++token.itemsConsumedFromCurrent == EXPLICIT_CONSUMER_CONSUMPTION_QUOTA_BEFORE_ROTATE) { + globalExplicitConsumerOffset.fetch_add(1, std::memory_order_relaxed); + } + return true; + } + + auto tail = producerListTail.load(std::memory_order_acquire); + auto ptr = static_cast(token.currentProducer)->next_prod(); + if (ptr == nullptr) { + ptr = tail; + } + while (ptr != static_cast(token.currentProducer)) { + if (ptr->dequeue(item)) { + token.currentProducer = ptr; + token.itemsConsumedFromCurrent = 1; + return true; + } + ptr = ptr->next_prod(); + if (ptr == nullptr) { + ptr = tail; + } + } + return false; + } + + // Attempts to dequeue several elements from the queue. + // Returns the number of items actually dequeued. + // Returns 0 if all producer streams appeared empty at the time they + // were checked (so, the queue is likely but not guaranteed to be empty). + // Never allocates. Thread-safe. + template + size_t try_dequeue_bulk(It itemFirst, size_t max) + { + size_t count = 0; + for (auto ptr = producerListTail.load(std::memory_order_acquire); ptr != nullptr; ptr = ptr->next_prod()) { + count += ptr->dequeue_bulk(itemFirst, max - count); + if (count == max) { + break; + } + } + return count; + } + + // Attempts to dequeue several elements from the queue using an explicit consumer token. + // Returns the number of items actually dequeued. + // Returns 0 if all producer streams appeared empty at the time they + // were checked (so, the queue is likely but not guaranteed to be empty). + // Never allocates. Thread-safe. + template + size_t try_dequeue_bulk(consumer_token_t& token, It itemFirst, size_t max) + { + if (token.desiredProducer == nullptr || token.lastKnownGlobalOffset != globalExplicitConsumerOffset.load(std::memory_order_relaxed)) { + if (!update_current_producer_after_rotation(token)) { + return 0; + } + } + + size_t count = static_cast(token.currentProducer)->dequeue_bulk(itemFirst, max); + if (count == max) { + if ((token.itemsConsumedFromCurrent += static_cast(max)) >= EXPLICIT_CONSUMER_CONSUMPTION_QUOTA_BEFORE_ROTATE) { + globalExplicitConsumerOffset.fetch_add(1, std::memory_order_relaxed); + } + return max; + } + token.itemsConsumedFromCurrent += static_cast(count); + max -= count; + + auto tail = producerListTail.load(std::memory_order_acquire); + auto ptr = static_cast(token.currentProducer)->next_prod(); + if (ptr == nullptr) { + ptr = tail; + } + while (ptr != static_cast(token.currentProducer)) { + auto dequeued = ptr->dequeue_bulk(itemFirst, max); + count += dequeued; + if (dequeued != 0) { + token.currentProducer = ptr; + token.itemsConsumedFromCurrent = static_cast(dequeued); + } + if (dequeued == max) { + break; + } + max -= dequeued; + ptr = ptr->next_prod(); + if (ptr == nullptr) { + ptr = tail; + } + } + return count; + } + + + + // Attempts to dequeue from a specific producer's inner queue. + // If you happen to know which producer you want to dequeue from, this + // is significantly faster than using the general-case try_dequeue methods. + // Returns false if the producer's queue appeared empty at the time it + // was checked (so, the queue is likely but not guaranteed to be empty). + // Never allocates. Thread-safe. + template + inline bool try_dequeue_from_producer(producer_token_t const& producer, U& item) + { + return static_cast(producer.producer)->dequeue(item); + } + + // Attempts to dequeue several elements from a specific producer's inner queue. + // Returns the number of items actually dequeued. + // If you happen to know which producer you want to dequeue from, this + // is significantly faster than using the general-case try_dequeue methods. + // Returns 0 if the producer's queue appeared empty at the time it + // was checked (so, the queue is likely but not guaranteed to be empty). + // Never allocates. Thread-safe. + template + inline size_t try_dequeue_bulk_from_producer(producer_token_t const& producer, It itemFirst, size_t max) + { + return static_cast(producer.producer)->dequeue_bulk(itemFirst, max); + } + + + // Returns an estimate of the total number of elements currently in the queue. This + // estimate is only accurate if the queue has completely stabilized before it is called + // (i.e. all enqueue and dequeue operations have completed and their memory effects are + // visible on the calling thread, and no further operations start while this method is + // being called). + // Thread-safe. + size_t size_approx() const + { + size_t size = 0; + for (auto ptr = producerListTail.load(std::memory_order_acquire); ptr != nullptr; ptr = ptr->next_prod()) { + size += ptr->size_approx(); + } + return size; + } + + + // Returns true if the underlying atomic variables used by + // the queue are lock-free (they should be on most platforms). + // Thread-safe. + static bool is_lock_free() + { + return + details::static_is_lock_free::value == 2 && + details::static_is_lock_free::value == 2 && + details::static_is_lock_free::value == 2 && + details::static_is_lock_free::value == 2 && + details::static_is_lock_free::value == 2 && + details::static_is_lock_free::thread_id_numeric_size_t>::value == 2; + } + + +private: + friend struct ProducerToken; + friend struct ConsumerToken; + struct ExplicitProducer; + friend struct ExplicitProducer; + struct ImplicitProducer; + friend struct ImplicitProducer; + friend class ConcurrentQueueTests; + + enum AllocationMode { CanAlloc, CannotAlloc }; + + + /////////////////////////////// + // Queue methods + /////////////////////////////// + + template + inline bool inner_enqueue(producer_token_t const& token, U&& element) + { + return static_cast(token.producer)->ConcurrentQueue::ExplicitProducer::template enqueue(std::forward(element)); + } + + template + inline bool inner_enqueue(U&& element) + { + auto producer = get_or_add_implicit_producer(); + return producer == nullptr ? false : producer->ConcurrentQueue::ImplicitProducer::template enqueue(std::forward(element)); + } + + template + inline bool inner_enqueue_bulk(producer_token_t const& token, It itemFirst, size_t count) + { + return static_cast(token.producer)->ConcurrentQueue::ExplicitProducer::template enqueue_bulk(itemFirst, count); + } + + template + inline bool inner_enqueue_bulk(It itemFirst, size_t count) + { + auto producer = get_or_add_implicit_producer(); + return producer == nullptr ? false : producer->ConcurrentQueue::ImplicitProducer::template enqueue_bulk(itemFirst, count); + } + + inline bool update_current_producer_after_rotation(consumer_token_t& token) + { + // Ah, there's been a rotation, figure out where we should be! + auto tail = producerListTail.load(std::memory_order_acquire); + if (token.desiredProducer == nullptr && tail == nullptr) { + return false; + } + auto prodCount = producerCount.load(std::memory_order_relaxed); + auto globalOffset = globalExplicitConsumerOffset.load(std::memory_order_relaxed); + if ((details::unlikely)(token.desiredProducer == nullptr)) { + // Aha, first time we're dequeueing anything. + // Figure out our local position + // Note: offset is from start, not end, but we're traversing from end -- subtract from count first + std::uint32_t offset = prodCount - 1 - (token.initialOffset % prodCount); + token.desiredProducer = tail; + for (std::uint32_t i = 0; i != offset; ++i) { + token.desiredProducer = static_cast(token.desiredProducer)->next_prod(); + if (token.desiredProducer == nullptr) { + token.desiredProducer = tail; + } + } + } + + std::uint32_t delta = globalOffset - token.lastKnownGlobalOffset; + if (delta >= prodCount) { + delta = delta % prodCount; + } + for (std::uint32_t i = 0; i != delta; ++i) { + token.desiredProducer = static_cast(token.desiredProducer)->next_prod(); + if (token.desiredProducer == nullptr) { + token.desiredProducer = tail; + } + } + + token.lastKnownGlobalOffset = globalOffset; + token.currentProducer = token.desiredProducer; + token.itemsConsumedFromCurrent = 0; + return true; + } + + + /////////////////////////// + // Free list + /////////////////////////// + + template + struct FreeListNode + { + FreeListNode() : freeListRefs(0), freeListNext(nullptr) { } + + std::atomic freeListRefs; + std::atomic freeListNext; + }; + + // A simple CAS-based lock-free free list. Not the fastest thing in the world under heavy contention, but + // simple and correct (assuming nodes are never freed until after the free list is destroyed), and fairly + // speedy under low contention. + template // N must inherit FreeListNode or have the same fields (and initialization of them) + struct FreeList + { + FreeList() : freeListHead(nullptr) { } + FreeList(FreeList&& other) : freeListHead(other.freeListHead.load(std::memory_order_relaxed)) { other.freeListHead.store(nullptr, std::memory_order_relaxed); } + void swap(FreeList& other) { details::swap_relaxed(freeListHead, other.freeListHead); } + + FreeList(FreeList const&) MOODYCAMEL_DELETE_FUNCTION; + FreeList& operator=(FreeList const&) MOODYCAMEL_DELETE_FUNCTION; + + inline void add(N* node) + { +#ifdef MCDBGQ_NOLOCKFREE_FREELIST + debug::DebugLock lock(mutex); +#endif + // We know that the should-be-on-freelist bit is 0 at this point, so it's safe to + // set it using a fetch_add + if (node->freeListRefs.fetch_add(SHOULD_BE_ON_FREELIST, std::memory_order_acq_rel) == 0) { + // Oh look! We were the last ones referencing this node, and we know + // we want to add it to the free list, so let's do it! + add_knowing_refcount_is_zero(node); + } + } + + inline N* try_get() + { +#ifdef MCDBGQ_NOLOCKFREE_FREELIST + debug::DebugLock lock(mutex); +#endif + auto head = freeListHead.load(std::memory_order_acquire); + while (head != nullptr) { + auto prevHead = head; + auto refs = head->freeListRefs.load(std::memory_order_relaxed); + if ((refs & REFS_MASK) == 0 || !head->freeListRefs.compare_exchange_strong(refs, refs + 1, std::memory_order_acquire, std::memory_order_relaxed)) { + head = freeListHead.load(std::memory_order_acquire); + continue; + } + + // Good, reference count has been incremented (it wasn't at zero), which means we can read the + // next and not worry about it changing between now and the time we do the CAS + auto next = head->freeListNext.load(std::memory_order_relaxed); + if (freeListHead.compare_exchange_strong(head, next, std::memory_order_acquire, std::memory_order_relaxed)) { + // Yay, got the node. This means it was on the list, which means shouldBeOnFreeList must be false no + // matter the refcount (because nobody else knows it's been taken off yet, it can't have been put back on). + assert((head->freeListRefs.load(std::memory_order_relaxed) & SHOULD_BE_ON_FREELIST) == 0); + + // Decrease refcount twice, once for our ref, and once for the list's ref + head->freeListRefs.fetch_sub(2, std::memory_order_release); + return head; + } + + // OK, the head must have changed on us, but we still need to decrease the refcount we increased. + // Note that we don't need to release any memory effects, but we do need to ensure that the reference + // count decrement happens-after the CAS on the head. + refs = prevHead->freeListRefs.fetch_sub(1, std::memory_order_acq_rel); + if (refs == SHOULD_BE_ON_FREELIST + 1) { + add_knowing_refcount_is_zero(prevHead); + } + } + + return nullptr; + } + + // Useful for traversing the list when there's no contention (e.g. to destroy remaining nodes) + N* head_unsafe() const { return freeListHead.load(std::memory_order_relaxed); } + + private: + inline void add_knowing_refcount_is_zero(N* node) + { + // Since the refcount is zero, and nobody can increase it once it's zero (except us, and we run + // only one copy of this method per node at a time, i.e. the single thread case), then we know + // we can safely change the next pointer of the node; however, once the refcount is back above + // zero, then other threads could increase it (happens under heavy contention, when the refcount + // goes to zero in between a load and a refcount increment of a node in try_get, then back up to + // something non-zero, then the refcount increment is done by the other thread) -- so, if the CAS + // to add the node to the actual list fails, decrease the refcount and leave the add operation to + // the next thread who puts the refcount back at zero (which could be us, hence the loop). + auto head = freeListHead.load(std::memory_order_relaxed); + while (true) { + node->freeListNext.store(head, std::memory_order_relaxed); + node->freeListRefs.store(1, std::memory_order_release); + if (!freeListHead.compare_exchange_strong(head, node, std::memory_order_release, std::memory_order_relaxed)) { + // Hmm, the add failed, but we can only try again when the refcount goes back to zero + if (node->freeListRefs.fetch_add(SHOULD_BE_ON_FREELIST - 1, std::memory_order_release) == 1) { + continue; + } + } + return; + } + } + + private: + // Implemented like a stack, but where node order doesn't matter (nodes are inserted out of order under contention) + std::atomic freeListHead; + + static const std::uint32_t REFS_MASK = 0x7FFFFFFF; + static const std::uint32_t SHOULD_BE_ON_FREELIST = 0x80000000; + +#ifdef MCDBGQ_NOLOCKFREE_FREELIST + debug::DebugMutex mutex; +#endif + }; + + + /////////////////////////// + // Block + /////////////////////////// + + enum InnerQueueContext { implicit_context = 0, explicit_context = 1 }; + + struct Block + { + Block() + : next(nullptr), elementsCompletelyDequeued(0), freeListRefs(0), freeListNext(nullptr), shouldBeOnFreeList(false), dynamicallyAllocated(true) + { +#ifdef MCDBGQ_TRACKMEM + owner = nullptr; +#endif + } + + template + inline bool is_empty() const + { + if (context == explicit_context && BLOCK_SIZE <= EXPLICIT_BLOCK_EMPTY_COUNTER_THRESHOLD) { + // Check flags + for (size_t i = 0; i < BLOCK_SIZE; ++i) { + if (!emptyFlags[i].load(std::memory_order_relaxed)) { + return false; + } + } + + // Aha, empty; make sure we have all other memory effects that happened before the empty flags were set + std::atomic_thread_fence(std::memory_order_acquire); + return true; + } + else { + // Check counter + if (elementsCompletelyDequeued.load(std::memory_order_relaxed) == BLOCK_SIZE) { + std::atomic_thread_fence(std::memory_order_acquire); + return true; + } + assert(elementsCompletelyDequeued.load(std::memory_order_relaxed) <= BLOCK_SIZE); + return false; + } + } + + // Returns true if the block is now empty (does not apply in explicit context) + template + inline bool set_empty(index_t i) + { + if (context == explicit_context && BLOCK_SIZE <= EXPLICIT_BLOCK_EMPTY_COUNTER_THRESHOLD) { + // Set flag + assert(!emptyFlags[BLOCK_SIZE - 1 - static_cast(i & static_cast(BLOCK_SIZE - 1))].load(std::memory_order_relaxed)); + emptyFlags[BLOCK_SIZE - 1 - static_cast(i & static_cast(BLOCK_SIZE - 1))].store(true, std::memory_order_release); + return false; + } + else { + // Increment counter + auto prevVal = elementsCompletelyDequeued.fetch_add(1, std::memory_order_release); + assert(prevVal < BLOCK_SIZE); + return prevVal == BLOCK_SIZE - 1; + } + } + + // Sets multiple contiguous item statuses to 'empty' (assumes no wrapping and count > 0). + // Returns true if the block is now empty (does not apply in explicit context). + template + inline bool set_many_empty(index_t i, size_t count) + { + if (context == explicit_context && BLOCK_SIZE <= EXPLICIT_BLOCK_EMPTY_COUNTER_THRESHOLD) { + // Set flags + std::atomic_thread_fence(std::memory_order_release); + i = BLOCK_SIZE - 1 - static_cast(i & static_cast(BLOCK_SIZE - 1)) - count + 1; + for (size_t j = 0; j != count; ++j) { + assert(!emptyFlags[i + j].load(std::memory_order_relaxed)); + emptyFlags[i + j].store(true, std::memory_order_relaxed); + } + return false; + } + else { + // Increment counter + auto prevVal = elementsCompletelyDequeued.fetch_add(count, std::memory_order_release); + assert(prevVal + count <= BLOCK_SIZE); + return prevVal + count == BLOCK_SIZE; + } + } + + template + inline void set_all_empty() + { + if (context == explicit_context && BLOCK_SIZE <= EXPLICIT_BLOCK_EMPTY_COUNTER_THRESHOLD) { + // Set all flags + for (size_t i = 0; i != BLOCK_SIZE; ++i) { + emptyFlags[i].store(true, std::memory_order_relaxed); + } + } + else { + // Reset counter + elementsCompletelyDequeued.store(BLOCK_SIZE, std::memory_order_relaxed); + } + } + + template + inline void reset_empty() + { + if (context == explicit_context && BLOCK_SIZE <= EXPLICIT_BLOCK_EMPTY_COUNTER_THRESHOLD) { + // Reset flags + for (size_t i = 0; i != BLOCK_SIZE; ++i) { + emptyFlags[i].store(false, std::memory_order_relaxed); + } + } + else { + // Reset counter + elementsCompletelyDequeued.store(0, std::memory_order_relaxed); + } + } + + inline T* operator[](index_t idx) MOODYCAMEL_NOEXCEPT { return static_cast(static_cast(elements)) + static_cast(idx & static_cast(BLOCK_SIZE - 1)); } + inline T const* operator[](index_t idx) const MOODYCAMEL_NOEXCEPT { return static_cast(static_cast(elements)) + static_cast(idx & static_cast(BLOCK_SIZE - 1)); } + + private: + // IMPORTANT: This must be the first member in Block, so that if T depends on the alignment of + // addresses returned by malloc, that alignment will be preserved. Apparently clang actually + // generates code that uses this assumption for AVX instructions in some cases. Ideally, we + // should also align Block to the alignment of T in case it's higher than malloc's 16-byte + // alignment, but this is hard to do in a cross-platform way. Assert for this case: + static_assert(std::alignment_of::value <= std::alignment_of::value, "The queue does not support super-aligned types at this time"); + // Additionally, we need the alignment of Block itself to be a multiple of max_align_t since + // otherwise the appropriate padding will not be added at the end of Block in order to make + // arrays of Blocks all be properly aligned (not just the first one). We use a union to force + // this. + union { + char elements[sizeof(T) * BLOCK_SIZE]; + details::max_align_t dummy; + }; + public: + Block* next; + std::atomic elementsCompletelyDequeued; + std::atomic emptyFlags[BLOCK_SIZE <= EXPLICIT_BLOCK_EMPTY_COUNTER_THRESHOLD ? BLOCK_SIZE : 1]; + public: + std::atomic freeListRefs; + std::atomic freeListNext; + std::atomic shouldBeOnFreeList; + bool dynamicallyAllocated; // Perhaps a better name for this would be 'isNotPartOfInitialBlockPool' + +#ifdef MCDBGQ_TRACKMEM + void* owner; +#endif + }; + static_assert(std::alignment_of::value >= std::alignment_of::value, "Internal error: Blocks must be at least as aligned as the type they are wrapping"); + + +#ifdef MCDBGQ_TRACKMEM +public: + struct MemStats; +private: +#endif + + /////////////////////////// + // Producer base + /////////////////////////// + + struct ProducerBase : public details::ConcurrentQueueProducerTypelessBase + { + ProducerBase(ConcurrentQueue* parent_, bool isExplicit_) : + tailIndex(0), + headIndex(0), + dequeueOptimisticCount(0), + dequeueOvercommit(0), + tailBlock(nullptr), + isExplicit(isExplicit_), + parent(parent_) + { + } + + virtual ~ProducerBase() { }; + + template + inline bool dequeue(U& element) + { + if (isExplicit) { + return static_cast(this)->dequeue(element); + } + else { + return static_cast(this)->dequeue(element); + } + } + + template + inline size_t dequeue_bulk(It& itemFirst, size_t max) + { + if (isExplicit) { + return static_cast(this)->dequeue_bulk(itemFirst, max); + } + else { + return static_cast(this)->dequeue_bulk(itemFirst, max); + } + } + + inline ProducerBase* next_prod() const { return static_cast(next); } + + inline size_t size_approx() const + { + auto tail = tailIndex.load(std::memory_order_relaxed); + auto head = headIndex.load(std::memory_order_relaxed); + return details::circular_less_than(head, tail) ? static_cast(tail - head) : 0; + } + + inline index_t getTail() const { return tailIndex.load(std::memory_order_relaxed); } + protected: + std::atomic tailIndex; // Where to enqueue to next + std::atomic headIndex; // Where to dequeue from next + + std::atomic dequeueOptimisticCount; + std::atomic dequeueOvercommit; + + Block* tailBlock; + + public: + bool isExplicit; + ConcurrentQueue* parent; + + protected: +#ifdef MCDBGQ_TRACKMEM + friend struct MemStats; +#endif + }; + + + /////////////////////////// + // Explicit queue + /////////////////////////// + + struct ExplicitProducer : public ProducerBase + { + explicit ExplicitProducer(ConcurrentQueue* parent) : + ProducerBase(parent, true), + blockIndex(nullptr), + pr_blockIndexSlotsUsed(0), + pr_blockIndexSize(EXPLICIT_INITIAL_INDEX_SIZE >> 1), + pr_blockIndexFront(0), + pr_blockIndexEntries(nullptr), + pr_blockIndexRaw(nullptr) + { + size_t poolBasedIndexSize = details::ceil_to_pow_2(parent->initialBlockPoolSize) >> 1; + if (poolBasedIndexSize > pr_blockIndexSize) { + pr_blockIndexSize = poolBasedIndexSize; + } + + new_block_index(0); // This creates an index with double the number of current entries, i.e. EXPLICIT_INITIAL_INDEX_SIZE + } + + ~ExplicitProducer() + { + // Destruct any elements not yet dequeued. + // Since we're in the destructor, we can assume all elements + // are either completely dequeued or completely not (no halfways). + if (this->tailBlock != nullptr) { // Note this means there must be a block index too + // First find the block that's partially dequeued, if any + Block* halfDequeuedBlock = nullptr; + if ((this->headIndex.load(std::memory_order_relaxed) & static_cast(BLOCK_SIZE - 1)) != 0) { + // The head's not on a block boundary, meaning a block somewhere is partially dequeued + // (or the head block is the tail block and was fully dequeued, but the head/tail are still not on a boundary) + size_t i = (pr_blockIndexFront - pr_blockIndexSlotsUsed) & (pr_blockIndexSize - 1); + while (details::circular_less_than(pr_blockIndexEntries[i].base + BLOCK_SIZE, this->headIndex.load(std::memory_order_relaxed))) { + i = (i + 1) & (pr_blockIndexSize - 1); + } + assert(details::circular_less_than(pr_blockIndexEntries[i].base, this->headIndex.load(std::memory_order_relaxed))); + halfDequeuedBlock = pr_blockIndexEntries[i].block; + } + + // Start at the head block (note the first line in the loop gives us the head from the tail on the first iteration) + auto block = this->tailBlock; + do { + block = block->next; + if (block->ConcurrentQueue::Block::template is_empty()) { + continue; + } + + size_t i = 0; // Offset into block + if (block == halfDequeuedBlock) { + i = static_cast(this->headIndex.load(std::memory_order_relaxed) & static_cast(BLOCK_SIZE - 1)); + } + + // Walk through all the items in the block; if this is the tail block, we need to stop when we reach the tail index + auto lastValidIndex = (this->tailIndex.load(std::memory_order_relaxed) & static_cast(BLOCK_SIZE - 1)) == 0 ? BLOCK_SIZE : static_cast(this->tailIndex.load(std::memory_order_relaxed) & static_cast(BLOCK_SIZE - 1)); + while (i != BLOCK_SIZE && (block != this->tailBlock || i != lastValidIndex)) { + (*block)[i++]->~T(); + } + } while (block != this->tailBlock); + } + + // Destroy all blocks that we own + if (this->tailBlock != nullptr) { + auto block = this->tailBlock; + do { + auto nextBlock = block->next; + if (block->dynamicallyAllocated) { + destroy(block); + } + else { + this->parent->add_block_to_free_list(block); + } + block = nextBlock; + } while (block != this->tailBlock); + } + + // Destroy the block indices + auto header = static_cast(pr_blockIndexRaw); + while (header != nullptr) { + auto prev = static_cast(header->prev); + header->~BlockIndexHeader(); + (Traits::free)(header); + header = prev; + } + } + + template + inline bool enqueue(U&& element) + { + index_t currentTailIndex = this->tailIndex.load(std::memory_order_relaxed); + index_t newTailIndex = 1 + currentTailIndex; + if ((currentTailIndex & static_cast(BLOCK_SIZE - 1)) == 0) { + // We reached the end of a block, start a new one + auto startBlock = this->tailBlock; + auto originalBlockIndexSlotsUsed = pr_blockIndexSlotsUsed; + if (this->tailBlock != nullptr && this->tailBlock->next->ConcurrentQueue::Block::template is_empty()) { + // We can re-use the block ahead of us, it's empty! + this->tailBlock = this->tailBlock->next; + this->tailBlock->ConcurrentQueue::Block::template reset_empty(); + + // We'll put the block on the block index (guaranteed to be room since we're conceptually removing the + // last block from it first -- except instead of removing then adding, we can just overwrite). + // Note that there must be a valid block index here, since even if allocation failed in the ctor, + // it would have been re-attempted when adding the first block to the queue; since there is such + // a block, a block index must have been successfully allocated. + } + else { + // Whatever head value we see here is >= the last value we saw here (relatively), + // and <= its current value. Since we have the most recent tail, the head must be + // <= to it. + auto head = this->headIndex.load(std::memory_order_relaxed); + assert(!details::circular_less_than(currentTailIndex, head)); + if (!details::circular_less_than(head, currentTailIndex + BLOCK_SIZE) + || (MAX_SUBQUEUE_SIZE != details::const_numeric_max::value && (MAX_SUBQUEUE_SIZE == 0 || MAX_SUBQUEUE_SIZE - BLOCK_SIZE < currentTailIndex - head))) { + // We can't enqueue in another block because there's not enough leeway -- the + // tail could surpass the head by the time the block fills up! (Or we'll exceed + // the size limit, if the second part of the condition was true.) + return false; + } + // We're going to need a new block; check that the block index has room + if (pr_blockIndexRaw == nullptr || pr_blockIndexSlotsUsed == pr_blockIndexSize) { + // Hmm, the circular block index is already full -- we'll need + // to allocate a new index. Note pr_blockIndexRaw can only be nullptr if + // the initial allocation failed in the constructor. + + if (allocMode == CannotAlloc || !new_block_index(pr_blockIndexSlotsUsed)) { + return false; + } + } + + // Insert a new block in the circular linked list + auto newBlock = this->parent->ConcurrentQueue::template requisition_block(); + if (newBlock == nullptr) { + return false; + } +#ifdef MCDBGQ_TRACKMEM + newBlock->owner = this; +#endif + newBlock->ConcurrentQueue::Block::template reset_empty(); + if (this->tailBlock == nullptr) { + newBlock->next = newBlock; + } + else { + newBlock->next = this->tailBlock->next; + this->tailBlock->next = newBlock; + } + this->tailBlock = newBlock; + ++pr_blockIndexSlotsUsed; + } + + if (!MOODYCAMEL_NOEXCEPT_CTOR(T, U, new ((T*)nullptr) T(std::forward(element)))) { + // The constructor may throw. We want the element not to appear in the queue in + // that case (without corrupting the queue): + MOODYCAMEL_TRY { + new ((*this->tailBlock)[currentTailIndex]) T(std::forward(element)); + } + MOODYCAMEL_CATCH (...) { + // Revert change to the current block, but leave the new block available + // for next time + pr_blockIndexSlotsUsed = originalBlockIndexSlotsUsed; + this->tailBlock = startBlock == nullptr ? this->tailBlock : startBlock; + MOODYCAMEL_RETHROW; + } + } + else { + (void)startBlock; + (void)originalBlockIndexSlotsUsed; + } + + // Add block to block index + auto& entry = blockIndex.load(std::memory_order_relaxed)->entries[pr_blockIndexFront]; + entry.base = currentTailIndex; + entry.block = this->tailBlock; + blockIndex.load(std::memory_order_relaxed)->front.store(pr_blockIndexFront, std::memory_order_release); + pr_blockIndexFront = (pr_blockIndexFront + 1) & (pr_blockIndexSize - 1); + + if (!MOODYCAMEL_NOEXCEPT_CTOR(T, U, new ((T*)nullptr) T(std::forward(element)))) { + this->tailIndex.store(newTailIndex, std::memory_order_release); + return true; + } + } + + // Enqueue + new ((*this->tailBlock)[currentTailIndex]) T(std::forward(element)); + + this->tailIndex.store(newTailIndex, std::memory_order_release); + return true; + } + + template + bool dequeue(U& element) + { + auto tail = this->tailIndex.load(std::memory_order_relaxed); + auto overcommit = this->dequeueOvercommit.load(std::memory_order_relaxed); + if (details::circular_less_than(this->dequeueOptimisticCount.load(std::memory_order_relaxed) - overcommit, tail)) { + // Might be something to dequeue, let's give it a try + + // Note that this if is purely for performance purposes in the common case when the queue is + // empty and the values are eventually consistent -- we may enter here spuriously. + + // Note that whatever the values of overcommit and tail are, they are not going to change (unless we + // change them) and must be the same value at this point (inside the if) as when the if condition was + // evaluated. + + // We insert an acquire fence here to synchronize-with the release upon incrementing dequeueOvercommit below. + // This ensures that whatever the value we got loaded into overcommit, the load of dequeueOptisticCount in + // the fetch_add below will result in a value at least as recent as that (and therefore at least as large). + // Note that I believe a compiler (signal) fence here would be sufficient due to the nature of fetch_add (all + // read-modify-write operations are guaranteed to work on the latest value in the modification order), but + // unfortunately that can't be shown to be correct using only the C++11 standard. + // See http://stackoverflow.com/questions/18223161/what-are-the-c11-memory-ordering-guarantees-in-this-corner-case + std::atomic_thread_fence(std::memory_order_acquire); + + // Increment optimistic counter, then check if it went over the boundary + auto myDequeueCount = this->dequeueOptimisticCount.fetch_add(1, std::memory_order_relaxed); + + // Note that since dequeueOvercommit must be <= dequeueOptimisticCount (because dequeueOvercommit is only ever + // incremented after dequeueOptimisticCount -- this is enforced in the `else` block below), and since we now + // have a version of dequeueOptimisticCount that is at least as recent as overcommit (due to the release upon + // incrementing dequeueOvercommit and the acquire above that synchronizes with it), overcommit <= myDequeueCount. + // However, we can't assert this since both dequeueOptimisticCount and dequeueOvercommit may (independently) + // overflow; in such a case, though, the logic still holds since the difference between the two is maintained. + + // Note that we reload tail here in case it changed; it will be the same value as before or greater, since + // this load is sequenced after (happens after) the earlier load above. This is supported by read-read + // coherency (as defined in the standard), explained here: http://en.cppreference.com/w/cpp/atomic/memory_order + tail = this->tailIndex.load(std::memory_order_acquire); + if ((details::likely)(details::circular_less_than(myDequeueCount - overcommit, tail))) { + // Guaranteed to be at least one element to dequeue! + + // Get the index. Note that since there's guaranteed to be at least one element, this + // will never exceed tail. We need to do an acquire-release fence here since it's possible + // that whatever condition got us to this point was for an earlier enqueued element (that + // we already see the memory effects for), but that by the time we increment somebody else + // has incremented it, and we need to see the memory effects for *that* element, which is + // in such a case is necessarily visible on the thread that incremented it in the first + // place with the more current condition (they must have acquired a tail that is at least + // as recent). + auto index = this->headIndex.fetch_add(1, std::memory_order_acq_rel); + + + // Determine which block the element is in + + auto localBlockIndex = blockIndex.load(std::memory_order_acquire); + auto localBlockIndexHead = localBlockIndex->front.load(std::memory_order_acquire); + + // We need to be careful here about subtracting and dividing because of index wrap-around. + // When an index wraps, we need to preserve the sign of the offset when dividing it by the + // block size (in order to get a correct signed block count offset in all cases): + auto headBase = localBlockIndex->entries[localBlockIndexHead].base; + auto blockBaseIndex = index & ~static_cast(BLOCK_SIZE - 1); + auto offset = static_cast(static_cast::type>(blockBaseIndex - headBase) / BLOCK_SIZE); + auto block = localBlockIndex->entries[(localBlockIndexHead + offset) & (localBlockIndex->size - 1)].block; + + // Dequeue + auto& el = *((*block)[index]); + if (!MOODYCAMEL_NOEXCEPT_ASSIGN(T, T&&, element = std::move(el))) { + // Make sure the element is still fully dequeued and destroyed even if the assignment + // throws + struct Guard { + Block* block; + index_t index; + + ~Guard() + { + (*block)[index]->~T(); + block->ConcurrentQueue::Block::template set_empty(index); + } + } guard = { block, index }; + + element = std::move(el); // NOLINT + } + else { + element = std::move(el); // NOLINT + el.~T(); // NOLINT + block->ConcurrentQueue::Block::template set_empty(index); + } + + return true; + } + else { + // Wasn't anything to dequeue after all; make the effective dequeue count eventually consistent + this->dequeueOvercommit.fetch_add(1, std::memory_order_release); // Release so that the fetch_add on dequeueOptimisticCount is guaranteed to happen before this write + } + } + + return false; + } + + template + bool enqueue_bulk(It itemFirst, size_t count) + { + // First, we need to make sure we have enough room to enqueue all of the elements; + // this means pre-allocating blocks and putting them in the block index (but only if + // all the allocations succeeded). + index_t startTailIndex = this->tailIndex.load(std::memory_order_relaxed); + auto startBlock = this->tailBlock; + auto originalBlockIndexFront = pr_blockIndexFront; + auto originalBlockIndexSlotsUsed = pr_blockIndexSlotsUsed; + + Block* firstAllocatedBlock = nullptr; + + // Figure out how many blocks we'll need to allocate, and do so + size_t blockBaseDiff = ((startTailIndex + count - 1) & ~static_cast(BLOCK_SIZE - 1)) - ((startTailIndex - 1) & ~static_cast(BLOCK_SIZE - 1)); + index_t currentTailIndex = (startTailIndex - 1) & ~static_cast(BLOCK_SIZE - 1); + if (blockBaseDiff > 0) { + // Allocate as many blocks as possible from ahead + while (blockBaseDiff > 0 && this->tailBlock != nullptr && this->tailBlock->next != firstAllocatedBlock && this->tailBlock->next->ConcurrentQueue::Block::template is_empty()) { + blockBaseDiff -= static_cast(BLOCK_SIZE); + currentTailIndex += static_cast(BLOCK_SIZE); + + this->tailBlock = this->tailBlock->next; + firstAllocatedBlock = firstAllocatedBlock == nullptr ? this->tailBlock : firstAllocatedBlock; + + auto& entry = blockIndex.load(std::memory_order_relaxed)->entries[pr_blockIndexFront]; + entry.base = currentTailIndex; + entry.block = this->tailBlock; + pr_blockIndexFront = (pr_blockIndexFront + 1) & (pr_blockIndexSize - 1); + } + + // Now allocate as many blocks as necessary from the block pool + while (blockBaseDiff > 0) { + blockBaseDiff -= static_cast(BLOCK_SIZE); + currentTailIndex += static_cast(BLOCK_SIZE); + + auto head = this->headIndex.load(std::memory_order_relaxed); + assert(!details::circular_less_than(currentTailIndex, head)); + bool full = !details::circular_less_than(head, currentTailIndex + BLOCK_SIZE) || (MAX_SUBQUEUE_SIZE != details::const_numeric_max::value && (MAX_SUBQUEUE_SIZE == 0 || MAX_SUBQUEUE_SIZE - BLOCK_SIZE < currentTailIndex - head)); + if (pr_blockIndexRaw == nullptr || pr_blockIndexSlotsUsed == pr_blockIndexSize || full) { + if (allocMode == CannotAlloc || full || !new_block_index(originalBlockIndexSlotsUsed)) { + // Failed to allocate, undo changes (but keep injected blocks) + pr_blockIndexFront = originalBlockIndexFront; + pr_blockIndexSlotsUsed = originalBlockIndexSlotsUsed; + this->tailBlock = startBlock == nullptr ? firstAllocatedBlock : startBlock; + return false; + } + + // pr_blockIndexFront is updated inside new_block_index, so we need to + // update our fallback value too (since we keep the new index even if we + // later fail) + originalBlockIndexFront = originalBlockIndexSlotsUsed; + } + + // Insert a new block in the circular linked list + auto newBlock = this->parent->ConcurrentQueue::template requisition_block(); + if (newBlock == nullptr) { + pr_blockIndexFront = originalBlockIndexFront; + pr_blockIndexSlotsUsed = originalBlockIndexSlotsUsed; + this->tailBlock = startBlock == nullptr ? firstAllocatedBlock : startBlock; + return false; + } + +#ifdef MCDBGQ_TRACKMEM + newBlock->owner = this; +#endif + newBlock->ConcurrentQueue::Block::template set_all_empty(); + if (this->tailBlock == nullptr) { + newBlock->next = newBlock; + } + else { + newBlock->next = this->tailBlock->next; + this->tailBlock->next = newBlock; + } + this->tailBlock = newBlock; + firstAllocatedBlock = firstAllocatedBlock == nullptr ? this->tailBlock : firstAllocatedBlock; + + ++pr_blockIndexSlotsUsed; + + auto& entry = blockIndex.load(std::memory_order_relaxed)->entries[pr_blockIndexFront]; + entry.base = currentTailIndex; + entry.block = this->tailBlock; + pr_blockIndexFront = (pr_blockIndexFront + 1) & (pr_blockIndexSize - 1); + } + + // Excellent, all allocations succeeded. Reset each block's emptiness before we fill them up, and + // publish the new block index front + auto block = firstAllocatedBlock; + while (true) { + block->ConcurrentQueue::Block::template reset_empty(); + if (block == this->tailBlock) { + break; + } + block = block->next; + } + + if (MOODYCAMEL_NOEXCEPT_CTOR(T, decltype(*itemFirst), new ((T*)nullptr) T(details::deref_noexcept(itemFirst)))) { + blockIndex.load(std::memory_order_relaxed)->front.store((pr_blockIndexFront - 1) & (pr_blockIndexSize - 1), std::memory_order_release); + } + } + + // Enqueue, one block at a time + index_t newTailIndex = startTailIndex + static_cast(count); + currentTailIndex = startTailIndex; + auto endBlock = this->tailBlock; + this->tailBlock = startBlock; + assert((startTailIndex & static_cast(BLOCK_SIZE - 1)) != 0 || firstAllocatedBlock != nullptr || count == 0); + if ((startTailIndex & static_cast(BLOCK_SIZE - 1)) == 0 && firstAllocatedBlock != nullptr) { + this->tailBlock = firstAllocatedBlock; + } + while (true) { + auto stopIndex = (currentTailIndex & ~static_cast(BLOCK_SIZE - 1)) + static_cast(BLOCK_SIZE); + if (details::circular_less_than(newTailIndex, stopIndex)) { + stopIndex = newTailIndex; + } + if (MOODYCAMEL_NOEXCEPT_CTOR(T, decltype(*itemFirst), new ((T*)nullptr) T(details::deref_noexcept(itemFirst)))) { + while (currentTailIndex != stopIndex) { + new ((*this->tailBlock)[currentTailIndex++]) T(*itemFirst++); + } + } + else { + MOODYCAMEL_TRY { + while (currentTailIndex != stopIndex) { + // Must use copy constructor even if move constructor is available + // because we may have to revert if there's an exception. + // Sorry about the horrible templated next line, but it was the only way + // to disable moving *at compile time*, which is important because a type + // may only define a (noexcept) move constructor, and so calls to the + // cctor will not compile, even if they are in an if branch that will never + // be executed + new ((*this->tailBlock)[currentTailIndex]) T(details::nomove_if<(bool)!MOODYCAMEL_NOEXCEPT_CTOR(T, decltype(*itemFirst), new ((T*)nullptr) T(details::deref_noexcept(itemFirst)))>::eval(*itemFirst)); + ++currentTailIndex; + ++itemFirst; + } + } + MOODYCAMEL_CATCH (...) { + // Oh dear, an exception's been thrown -- destroy the elements that + // were enqueued so far and revert the entire bulk operation (we'll keep + // any allocated blocks in our linked list for later, though). + auto constructedStopIndex = currentTailIndex; + auto lastBlockEnqueued = this->tailBlock; + + pr_blockIndexFront = originalBlockIndexFront; + pr_blockIndexSlotsUsed = originalBlockIndexSlotsUsed; + this->tailBlock = startBlock == nullptr ? firstAllocatedBlock : startBlock; + + if (!details::is_trivially_destructible::value) { + auto block = startBlock; + if ((startTailIndex & static_cast(BLOCK_SIZE - 1)) == 0) { + block = firstAllocatedBlock; + } + currentTailIndex = startTailIndex; + while (true) { + stopIndex = (currentTailIndex & ~static_cast(BLOCK_SIZE - 1)) + static_cast(BLOCK_SIZE); + if (details::circular_less_than(constructedStopIndex, stopIndex)) { + stopIndex = constructedStopIndex; + } + while (currentTailIndex != stopIndex) { + (*block)[currentTailIndex++]->~T(); + } + if (block == lastBlockEnqueued) { + break; + } + block = block->next; + } + } + MOODYCAMEL_RETHROW; + } + } + + if (this->tailBlock == endBlock) { + assert(currentTailIndex == newTailIndex); + break; + } + this->tailBlock = this->tailBlock->next; + } + + if (!MOODYCAMEL_NOEXCEPT_CTOR(T, decltype(*itemFirst), new ((T*)nullptr) T(details::deref_noexcept(itemFirst))) && firstAllocatedBlock != nullptr) { + blockIndex.load(std::memory_order_relaxed)->front.store((pr_blockIndexFront - 1) & (pr_blockIndexSize - 1), std::memory_order_release); + } + + this->tailIndex.store(newTailIndex, std::memory_order_release); + return true; + } + + template + size_t dequeue_bulk(It& itemFirst, size_t max) + { + auto tail = this->tailIndex.load(std::memory_order_relaxed); + auto overcommit = this->dequeueOvercommit.load(std::memory_order_relaxed); + auto desiredCount = static_cast(tail - (this->dequeueOptimisticCount.load(std::memory_order_relaxed) - overcommit)); + if (details::circular_less_than(0, desiredCount)) { + desiredCount = desiredCount < max ? desiredCount : max; + std::atomic_thread_fence(std::memory_order_acquire); + + auto myDequeueCount = this->dequeueOptimisticCount.fetch_add(desiredCount, std::memory_order_relaxed);; + + tail = this->tailIndex.load(std::memory_order_acquire); + auto actualCount = static_cast(tail - (myDequeueCount - overcommit)); + if (details::circular_less_than(0, actualCount)) { + actualCount = desiredCount < actualCount ? desiredCount : actualCount; + if (actualCount < desiredCount) { + this->dequeueOvercommit.fetch_add(desiredCount - actualCount, std::memory_order_release); + } + + // Get the first index. Note that since there's guaranteed to be at least actualCount elements, this + // will never exceed tail. + auto firstIndex = this->headIndex.fetch_add(actualCount, std::memory_order_acq_rel); + + // Determine which block the first element is in + auto localBlockIndex = blockIndex.load(std::memory_order_acquire); + auto localBlockIndexHead = localBlockIndex->front.load(std::memory_order_acquire); + + auto headBase = localBlockIndex->entries[localBlockIndexHead].base; + auto firstBlockBaseIndex = firstIndex & ~static_cast(BLOCK_SIZE - 1); + auto offset = static_cast(static_cast::type>(firstBlockBaseIndex - headBase) / BLOCK_SIZE); + auto indexIndex = (localBlockIndexHead + offset) & (localBlockIndex->size - 1); + + // Iterate the blocks and dequeue + auto index = firstIndex; + do { + auto firstIndexInBlock = index; + auto endIndex = (index & ~static_cast(BLOCK_SIZE - 1)) + static_cast(BLOCK_SIZE); + endIndex = details::circular_less_than(firstIndex + static_cast(actualCount), endIndex) ? firstIndex + static_cast(actualCount) : endIndex; + auto block = localBlockIndex->entries[indexIndex].block; + if (MOODYCAMEL_NOEXCEPT_ASSIGN(T, T&&, details::deref_noexcept(itemFirst) = std::move((*(*block)[index])))) { + while (index != endIndex) { + auto& el = *((*block)[index]); + *itemFirst++ = std::move(el); + el.~T(); + ++index; + } + } + else { + MOODYCAMEL_TRY { + while (index != endIndex) { + auto& el = *((*block)[index]); + *itemFirst = std::move(el); + ++itemFirst; + el.~T(); + ++index; + } + } + MOODYCAMEL_CATCH (...) { + // It's too late to revert the dequeue, but we can make sure that all + // the dequeued objects are properly destroyed and the block index + // (and empty count) are properly updated before we propagate the exception + do { + block = localBlockIndex->entries[indexIndex].block; + while (index != endIndex) { + (*block)[index++]->~T(); + } + block->ConcurrentQueue::Block::template set_many_empty(firstIndexInBlock, static_cast(endIndex - firstIndexInBlock)); + indexIndex = (indexIndex + 1) & (localBlockIndex->size - 1); + + firstIndexInBlock = index; + endIndex = (index & ~static_cast(BLOCK_SIZE - 1)) + static_cast(BLOCK_SIZE); + endIndex = details::circular_less_than(firstIndex + static_cast(actualCount), endIndex) ? firstIndex + static_cast(actualCount) : endIndex; + } while (index != firstIndex + actualCount); + + MOODYCAMEL_RETHROW; + } + } + block->ConcurrentQueue::Block::template set_many_empty(firstIndexInBlock, static_cast(endIndex - firstIndexInBlock)); + indexIndex = (indexIndex + 1) & (localBlockIndex->size - 1); + } while (index != firstIndex + actualCount); + + return actualCount; + } + else { + // Wasn't anything to dequeue after all; make the effective dequeue count eventually consistent + this->dequeueOvercommit.fetch_add(desiredCount, std::memory_order_release); + } + } + + return 0; + } + + private: + struct BlockIndexEntry + { + index_t base; + Block* block; + }; + + struct BlockIndexHeader + { + size_t size; + std::atomic front; // Current slot (not next, like pr_blockIndexFront) + BlockIndexEntry* entries; + void* prev; + }; + + + bool new_block_index(size_t numberOfFilledSlotsToExpose) + { + auto prevBlockSizeMask = pr_blockIndexSize - 1; + + // Create the new block + pr_blockIndexSize <<= 1; + auto newRawPtr = static_cast((Traits::malloc)(sizeof(BlockIndexHeader) + std::alignment_of::value - 1 + sizeof(BlockIndexEntry) * pr_blockIndexSize)); + if (newRawPtr == nullptr) { + pr_blockIndexSize >>= 1; // Reset to allow graceful retry + return false; + } + + auto newBlockIndexEntries = reinterpret_cast(details::align_for(newRawPtr + sizeof(BlockIndexHeader))); + + // Copy in all the old indices, if any + size_t j = 0; + if (pr_blockIndexSlotsUsed != 0) { + auto i = (pr_blockIndexFront - pr_blockIndexSlotsUsed) & prevBlockSizeMask; + do { + newBlockIndexEntries[j++] = pr_blockIndexEntries[i]; + i = (i + 1) & prevBlockSizeMask; + } while (i != pr_blockIndexFront); + } + + // Update everything + auto header = new (newRawPtr) BlockIndexHeader; + header->size = pr_blockIndexSize; + header->front.store(numberOfFilledSlotsToExpose - 1, std::memory_order_relaxed); + header->entries = newBlockIndexEntries; + header->prev = pr_blockIndexRaw; // we link the new block to the old one so we can free it later + + pr_blockIndexFront = j; + pr_blockIndexEntries = newBlockIndexEntries; + pr_blockIndexRaw = newRawPtr; + blockIndex.store(header, std::memory_order_release); + + return true; + } + + private: + std::atomic blockIndex; + + // To be used by producer only -- consumer must use the ones in referenced by blockIndex + size_t pr_blockIndexSlotsUsed; + size_t pr_blockIndexSize; + size_t pr_blockIndexFront; // Next slot (not current) + BlockIndexEntry* pr_blockIndexEntries; + void* pr_blockIndexRaw; + +#ifdef MOODYCAMEL_QUEUE_INTERNAL_DEBUG + public: + ExplicitProducer* nextExplicitProducer; + private: +#endif + +#ifdef MCDBGQ_TRACKMEM + friend struct MemStats; +#endif + }; + + + ////////////////////////////////// + // Implicit queue + ////////////////////////////////// + + struct ImplicitProducer : public ProducerBase + { + ImplicitProducer(ConcurrentQueue* parent) : + ProducerBase(parent, false), + nextBlockIndexCapacity(IMPLICIT_INITIAL_INDEX_SIZE), + blockIndex(nullptr) + { + new_block_index(); + } + + ~ImplicitProducer() + { + // Note that since we're in the destructor we can assume that all enqueue/dequeue operations + // completed already; this means that all undequeued elements are placed contiguously across + // contiguous blocks, and that only the first and last remaining blocks can be only partially + // empty (all other remaining blocks must be completely full). + +#ifdef MOODYCAMEL_CPP11_THREAD_LOCAL_SUPPORTED + // Unregister ourselves for thread termination notification + if (!this->inactive.load(std::memory_order_relaxed)) { + details::ThreadExitNotifier::unsubscribe(&threadExitListener); + } +#endif + + // Destroy all remaining elements! + auto tail = this->tailIndex.load(std::memory_order_relaxed); + auto index = this->headIndex.load(std::memory_order_relaxed); + Block* block = nullptr; + assert(index == tail || details::circular_less_than(index, tail)); + bool forceFreeLastBlock = index != tail; // If we enter the loop, then the last (tail) block will not be freed + while (index != tail) { + if ((index & static_cast(BLOCK_SIZE - 1)) == 0 || block == nullptr) { + if (block != nullptr) { + // Free the old block + this->parent->add_block_to_free_list(block); + } + + block = get_block_index_entry_for_index(index)->value.load(std::memory_order_relaxed); + } + + ((*block)[index])->~T(); + ++index; + } + // Even if the queue is empty, there's still one block that's not on the free list + // (unless the head index reached the end of it, in which case the tail will be poised + // to create a new block). + if (this->tailBlock != nullptr && (forceFreeLastBlock || (tail & static_cast(BLOCK_SIZE - 1)) != 0)) { + this->parent->add_block_to_free_list(this->tailBlock); + } + + // Destroy block index + auto localBlockIndex = blockIndex.load(std::memory_order_relaxed); + if (localBlockIndex != nullptr) { + for (size_t i = 0; i != localBlockIndex->capacity; ++i) { + localBlockIndex->index[i]->~BlockIndexEntry(); + } + do { + auto prev = localBlockIndex->prev; + localBlockIndex->~BlockIndexHeader(); + (Traits::free)(localBlockIndex); + localBlockIndex = prev; + } while (localBlockIndex != nullptr); + } + } + + template + inline bool enqueue(U&& element) + { + index_t currentTailIndex = this->tailIndex.load(std::memory_order_relaxed); + index_t newTailIndex = 1 + currentTailIndex; + if ((currentTailIndex & static_cast(BLOCK_SIZE - 1)) == 0) { + // We reached the end of a block, start a new one + auto head = this->headIndex.load(std::memory_order_relaxed); + assert(!details::circular_less_than(currentTailIndex, head)); + if (!details::circular_less_than(head, currentTailIndex + BLOCK_SIZE) || (MAX_SUBQUEUE_SIZE != details::const_numeric_max::value && (MAX_SUBQUEUE_SIZE == 0 || MAX_SUBQUEUE_SIZE - BLOCK_SIZE < currentTailIndex - head))) { + return false; + } +#ifdef MCDBGQ_NOLOCKFREE_IMPLICITPRODBLOCKINDEX + debug::DebugLock lock(mutex); +#endif + // Find out where we'll be inserting this block in the block index + BlockIndexEntry* idxEntry; + if (!insert_block_index_entry(idxEntry, currentTailIndex)) { + return false; + } + + // Get ahold of a new block + auto newBlock = this->parent->ConcurrentQueue::template requisition_block(); + if (newBlock == nullptr) { + rewind_block_index_tail(); + idxEntry->value.store(nullptr, std::memory_order_relaxed); + return false; + } +#ifdef MCDBGQ_TRACKMEM + newBlock->owner = this; +#endif + newBlock->ConcurrentQueue::Block::template reset_empty(); + + if (!MOODYCAMEL_NOEXCEPT_CTOR(T, U, new ((T*)nullptr) T(std::forward(element)))) { + // May throw, try to insert now before we publish the fact that we have this new block + MOODYCAMEL_TRY { + new ((*newBlock)[currentTailIndex]) T(std::forward(element)); + } + MOODYCAMEL_CATCH (...) { + rewind_block_index_tail(); + idxEntry->value.store(nullptr, std::memory_order_relaxed); + this->parent->add_block_to_free_list(newBlock); + MOODYCAMEL_RETHROW; + } + } + + // Insert the new block into the index + idxEntry->value.store(newBlock, std::memory_order_relaxed); + + this->tailBlock = newBlock; + + if (!MOODYCAMEL_NOEXCEPT_CTOR(T, U, new ((T*)nullptr) T(std::forward(element)))) { + this->tailIndex.store(newTailIndex, std::memory_order_release); + return true; + } + } + + // Enqueue + new ((*this->tailBlock)[currentTailIndex]) T(std::forward(element)); + + this->tailIndex.store(newTailIndex, std::memory_order_release); + return true; + } + + template + bool dequeue(U& element) + { + // See ExplicitProducer::dequeue for rationale and explanation + index_t tail = this->tailIndex.load(std::memory_order_relaxed); + index_t overcommit = this->dequeueOvercommit.load(std::memory_order_relaxed); + if (details::circular_less_than(this->dequeueOptimisticCount.load(std::memory_order_relaxed) - overcommit, tail)) { + std::atomic_thread_fence(std::memory_order_acquire); + + index_t myDequeueCount = this->dequeueOptimisticCount.fetch_add(1, std::memory_order_relaxed); + tail = this->tailIndex.load(std::memory_order_acquire); + if ((details::likely)(details::circular_less_than(myDequeueCount - overcommit, tail))) { + index_t index = this->headIndex.fetch_add(1, std::memory_order_acq_rel); + + // Determine which block the element is in + auto entry = get_block_index_entry_for_index(index); + + // Dequeue + auto block = entry->value.load(std::memory_order_relaxed); + auto& el = *((*block)[index]); + + if (!MOODYCAMEL_NOEXCEPT_ASSIGN(T, T&&, element = std::move(el))) { +#ifdef MCDBGQ_NOLOCKFREE_IMPLICITPRODBLOCKINDEX + // Note: Acquiring the mutex with every dequeue instead of only when a block + // is released is very sub-optimal, but it is, after all, purely debug code. + debug::DebugLock lock(producer->mutex); +#endif + struct Guard { + Block* block; + index_t index; + BlockIndexEntry* entry; + ConcurrentQueue* parent; + + ~Guard() + { + (*block)[index]->~T(); + if (block->ConcurrentQueue::Block::template set_empty(index)) { + entry->value.store(nullptr, std::memory_order_relaxed); + parent->add_block_to_free_list(block); + } + } + } guard = { block, index, entry, this->parent }; + + element = std::move(el); // NOLINT + } + else { + element = std::move(el); // NOLINT + el.~T(); // NOLINT + + if (block->ConcurrentQueue::Block::template set_empty(index)) { + { +#ifdef MCDBGQ_NOLOCKFREE_IMPLICITPRODBLOCKINDEX + debug::DebugLock lock(mutex); +#endif + // Add the block back into the global free pool (and remove from block index) + entry->value.store(nullptr, std::memory_order_relaxed); + } + this->parent->add_block_to_free_list(block); // releases the above store + } + } + + return true; + } + else { + this->dequeueOvercommit.fetch_add(1, std::memory_order_release); + } + } + + return false; + } + + template + bool enqueue_bulk(It itemFirst, size_t count) + { + // First, we need to make sure we have enough room to enqueue all of the elements; + // this means pre-allocating blocks and putting them in the block index (but only if + // all the allocations succeeded). + + // Note that the tailBlock we start off with may not be owned by us any more; + // this happens if it was filled up exactly to the top (setting tailIndex to + // the first index of the next block which is not yet allocated), then dequeued + // completely (putting it on the free list) before we enqueue again. + + index_t startTailIndex = this->tailIndex.load(std::memory_order_relaxed); + auto startBlock = this->tailBlock; + Block* firstAllocatedBlock = nullptr; + auto endBlock = this->tailBlock; + + // Figure out how many blocks we'll need to allocate, and do so + size_t blockBaseDiff = ((startTailIndex + count - 1) & ~static_cast(BLOCK_SIZE - 1)) - ((startTailIndex - 1) & ~static_cast(BLOCK_SIZE - 1)); + index_t currentTailIndex = (startTailIndex - 1) & ~static_cast(BLOCK_SIZE - 1); + if (blockBaseDiff > 0) { +#ifdef MCDBGQ_NOLOCKFREE_IMPLICITPRODBLOCKINDEX + debug::DebugLock lock(mutex); +#endif + do { + blockBaseDiff -= static_cast(BLOCK_SIZE); + currentTailIndex += static_cast(BLOCK_SIZE); + + // Find out where we'll be inserting this block in the block index + BlockIndexEntry* idxEntry = nullptr; // initialization here unnecessary but compiler can't always tell + Block* newBlock; + bool indexInserted = false; + auto head = this->headIndex.load(std::memory_order_relaxed); + assert(!details::circular_less_than(currentTailIndex, head)); + bool full = !details::circular_less_than(head, currentTailIndex + BLOCK_SIZE) || (MAX_SUBQUEUE_SIZE != details::const_numeric_max::value && (MAX_SUBQUEUE_SIZE == 0 || MAX_SUBQUEUE_SIZE - BLOCK_SIZE < currentTailIndex - head)); + if (full || !(indexInserted = insert_block_index_entry(idxEntry, currentTailIndex)) || (newBlock = this->parent->ConcurrentQueue::template requisition_block()) == nullptr) { + // Index allocation or block allocation failed; revert any other allocations + // and index insertions done so far for this operation + if (indexInserted) { + rewind_block_index_tail(); + idxEntry->value.store(nullptr, std::memory_order_relaxed); + } + currentTailIndex = (startTailIndex - 1) & ~static_cast(BLOCK_SIZE - 1); + for (auto block = firstAllocatedBlock; block != nullptr; block = block->next) { + currentTailIndex += static_cast(BLOCK_SIZE); + idxEntry = get_block_index_entry_for_index(currentTailIndex); + idxEntry->value.store(nullptr, std::memory_order_relaxed); + rewind_block_index_tail(); + } + this->parent->add_blocks_to_free_list(firstAllocatedBlock); + this->tailBlock = startBlock; + + return false; + } + +#ifdef MCDBGQ_TRACKMEM + newBlock->owner = this; +#endif + newBlock->ConcurrentQueue::Block::template reset_empty(); + newBlock->next = nullptr; + + // Insert the new block into the index + idxEntry->value.store(newBlock, std::memory_order_relaxed); + + // Store the chain of blocks so that we can undo if later allocations fail, + // and so that we can find the blocks when we do the actual enqueueing + if ((startTailIndex & static_cast(BLOCK_SIZE - 1)) != 0 || firstAllocatedBlock != nullptr) { + assert(this->tailBlock != nullptr); + this->tailBlock->next = newBlock; + } + this->tailBlock = newBlock; + endBlock = newBlock; + firstAllocatedBlock = firstAllocatedBlock == nullptr ? newBlock : firstAllocatedBlock; + } while (blockBaseDiff > 0); + } + + // Enqueue, one block at a time + index_t newTailIndex = startTailIndex + static_cast(count); + currentTailIndex = startTailIndex; + this->tailBlock = startBlock; + assert((startTailIndex & static_cast(BLOCK_SIZE - 1)) != 0 || firstAllocatedBlock != nullptr || count == 0); + if ((startTailIndex & static_cast(BLOCK_SIZE - 1)) == 0 && firstAllocatedBlock != nullptr) { + this->tailBlock = firstAllocatedBlock; + } + while (true) { + auto stopIndex = (currentTailIndex & ~static_cast(BLOCK_SIZE - 1)) + static_cast(BLOCK_SIZE); + if (details::circular_less_than(newTailIndex, stopIndex)) { + stopIndex = newTailIndex; + } + if (MOODYCAMEL_NOEXCEPT_CTOR(T, decltype(*itemFirst), new ((T*)nullptr) T(details::deref_noexcept(itemFirst)))) { + while (currentTailIndex != stopIndex) { + new ((*this->tailBlock)[currentTailIndex++]) T(*itemFirst++); + } + } + else { + MOODYCAMEL_TRY { + while (currentTailIndex != stopIndex) { + new ((*this->tailBlock)[currentTailIndex]) T(details::nomove_if<(bool)!MOODYCAMEL_NOEXCEPT_CTOR(T, decltype(*itemFirst), new ((T*)nullptr) T(details::deref_noexcept(itemFirst)))>::eval(*itemFirst)); + ++currentTailIndex; + ++itemFirst; + } + } + MOODYCAMEL_CATCH (...) { + auto constructedStopIndex = currentTailIndex; + auto lastBlockEnqueued = this->tailBlock; + + if (!details::is_trivially_destructible::value) { + auto block = startBlock; + if ((startTailIndex & static_cast(BLOCK_SIZE - 1)) == 0) { + block = firstAllocatedBlock; + } + currentTailIndex = startTailIndex; + while (true) { + stopIndex = (currentTailIndex & ~static_cast(BLOCK_SIZE - 1)) + static_cast(BLOCK_SIZE); + if (details::circular_less_than(constructedStopIndex, stopIndex)) { + stopIndex = constructedStopIndex; + } + while (currentTailIndex != stopIndex) { + (*block)[currentTailIndex++]->~T(); + } + if (block == lastBlockEnqueued) { + break; + } + block = block->next; + } + } + + currentTailIndex = (startTailIndex - 1) & ~static_cast(BLOCK_SIZE - 1); + for (auto block = firstAllocatedBlock; block != nullptr; block = block->next) { + currentTailIndex += static_cast(BLOCK_SIZE); + auto idxEntry = get_block_index_entry_for_index(currentTailIndex); + idxEntry->value.store(nullptr, std::memory_order_relaxed); + rewind_block_index_tail(); + } + this->parent->add_blocks_to_free_list(firstAllocatedBlock); + this->tailBlock = startBlock; + MOODYCAMEL_RETHROW; + } + } + + if (this->tailBlock == endBlock) { + assert(currentTailIndex == newTailIndex); + break; + } + this->tailBlock = this->tailBlock->next; + } + this->tailIndex.store(newTailIndex, std::memory_order_release); + return true; + } + + template + size_t dequeue_bulk(It& itemFirst, size_t max) + { + auto tail = this->tailIndex.load(std::memory_order_relaxed); + auto overcommit = this->dequeueOvercommit.load(std::memory_order_relaxed); + auto desiredCount = static_cast(tail - (this->dequeueOptimisticCount.load(std::memory_order_relaxed) - overcommit)); + if (details::circular_less_than(0, desiredCount)) { + desiredCount = desiredCount < max ? desiredCount : max; + std::atomic_thread_fence(std::memory_order_acquire); + + auto myDequeueCount = this->dequeueOptimisticCount.fetch_add(desiredCount, std::memory_order_relaxed); + + tail = this->tailIndex.load(std::memory_order_acquire); + auto actualCount = static_cast(tail - (myDequeueCount - overcommit)); + if (details::circular_less_than(0, actualCount)) { + actualCount = desiredCount < actualCount ? desiredCount : actualCount; + if (actualCount < desiredCount) { + this->dequeueOvercommit.fetch_add(desiredCount - actualCount, std::memory_order_release); + } + + // Get the first index. Note that since there's guaranteed to be at least actualCount elements, this + // will never exceed tail. + auto firstIndex = this->headIndex.fetch_add(actualCount, std::memory_order_acq_rel); + + // Iterate the blocks and dequeue + auto index = firstIndex; + BlockIndexHeader* localBlockIndex; + auto indexIndex = get_block_index_index_for_index(index, localBlockIndex); + do { + auto blockStartIndex = index; + auto endIndex = (index & ~static_cast(BLOCK_SIZE - 1)) + static_cast(BLOCK_SIZE); + endIndex = details::circular_less_than(firstIndex + static_cast(actualCount), endIndex) ? firstIndex + static_cast(actualCount) : endIndex; + + auto entry = localBlockIndex->index[indexIndex]; + auto block = entry->value.load(std::memory_order_relaxed); + if (MOODYCAMEL_NOEXCEPT_ASSIGN(T, T&&, details::deref_noexcept(itemFirst) = std::move((*(*block)[index])))) { + while (index != endIndex) { + auto& el = *((*block)[index]); + *itemFirst++ = std::move(el); + el.~T(); + ++index; + } + } + else { + MOODYCAMEL_TRY { + while (index != endIndex) { + auto& el = *((*block)[index]); + *itemFirst = std::move(el); + ++itemFirst; + el.~T(); + ++index; + } + } + MOODYCAMEL_CATCH (...) { + do { + entry = localBlockIndex->index[indexIndex]; + block = entry->value.load(std::memory_order_relaxed); + while (index != endIndex) { + (*block)[index++]->~T(); + } + + if (block->ConcurrentQueue::Block::template set_many_empty(blockStartIndex, static_cast(endIndex - blockStartIndex))) { +#ifdef MCDBGQ_NOLOCKFREE_IMPLICITPRODBLOCKINDEX + debug::DebugLock lock(mutex); +#endif + entry->value.store(nullptr, std::memory_order_relaxed); + this->parent->add_block_to_free_list(block); + } + indexIndex = (indexIndex + 1) & (localBlockIndex->capacity - 1); + + blockStartIndex = index; + endIndex = (index & ~static_cast(BLOCK_SIZE - 1)) + static_cast(BLOCK_SIZE); + endIndex = details::circular_less_than(firstIndex + static_cast(actualCount), endIndex) ? firstIndex + static_cast(actualCount) : endIndex; + } while (index != firstIndex + actualCount); + + MOODYCAMEL_RETHROW; + } + } + if (block->ConcurrentQueue::Block::template set_many_empty(blockStartIndex, static_cast(endIndex - blockStartIndex))) { + { +#ifdef MCDBGQ_NOLOCKFREE_IMPLICITPRODBLOCKINDEX + debug::DebugLock lock(mutex); +#endif + // Note that the set_many_empty above did a release, meaning that anybody who acquires the block + // we're about to free can use it safely since our writes (and reads!) will have happened-before then. + entry->value.store(nullptr, std::memory_order_relaxed); + } + this->parent->add_block_to_free_list(block); // releases the above store + } + indexIndex = (indexIndex + 1) & (localBlockIndex->capacity - 1); + } while (index != firstIndex + actualCount); + + return actualCount; + } + else { + this->dequeueOvercommit.fetch_add(desiredCount, std::memory_order_release); + } + } + + return 0; + } + + private: + // The block size must be > 1, so any number with the low bit set is an invalid block base index + static const index_t INVALID_BLOCK_BASE = 1; + + struct BlockIndexEntry + { + std::atomic key; + std::atomic value; + }; + + struct BlockIndexHeader + { + size_t capacity; + std::atomic tail; + BlockIndexEntry* entries; + BlockIndexEntry** index; + BlockIndexHeader* prev; + }; + + template + inline bool insert_block_index_entry(BlockIndexEntry*& idxEntry, index_t blockStartIndex) + { + auto localBlockIndex = blockIndex.load(std::memory_order_relaxed); // We're the only writer thread, relaxed is OK + if (localBlockIndex == nullptr) { + return false; // this can happen if new_block_index failed in the constructor + } + auto newTail = (localBlockIndex->tail.load(std::memory_order_relaxed) + 1) & (localBlockIndex->capacity - 1); + idxEntry = localBlockIndex->index[newTail]; + if (idxEntry->key.load(std::memory_order_relaxed) == INVALID_BLOCK_BASE || + idxEntry->value.load(std::memory_order_relaxed) == nullptr) { + + idxEntry->key.store(blockStartIndex, std::memory_order_relaxed); + localBlockIndex->tail.store(newTail, std::memory_order_release); + return true; + } + + // No room in the old block index, try to allocate another one! + if (allocMode == CannotAlloc || !new_block_index()) { + return false; + } + localBlockIndex = blockIndex.load(std::memory_order_relaxed); + newTail = (localBlockIndex->tail.load(std::memory_order_relaxed) + 1) & (localBlockIndex->capacity - 1); + idxEntry = localBlockIndex->index[newTail]; + assert(idxEntry->key.load(std::memory_order_relaxed) == INVALID_BLOCK_BASE); + idxEntry->key.store(blockStartIndex, std::memory_order_relaxed); + localBlockIndex->tail.store(newTail, std::memory_order_release); + return true; + } + + inline void rewind_block_index_tail() + { + auto localBlockIndex = blockIndex.load(std::memory_order_relaxed); + localBlockIndex->tail.store((localBlockIndex->tail.load(std::memory_order_relaxed) - 1) & (localBlockIndex->capacity - 1), std::memory_order_relaxed); + } + + inline BlockIndexEntry* get_block_index_entry_for_index(index_t index) const + { + BlockIndexHeader* localBlockIndex; + auto idx = get_block_index_index_for_index(index, localBlockIndex); + return localBlockIndex->index[idx]; + } + + inline size_t get_block_index_index_for_index(index_t index, BlockIndexHeader*& localBlockIndex) const + { +#ifdef MCDBGQ_NOLOCKFREE_IMPLICITPRODBLOCKINDEX + debug::DebugLock lock(mutex); +#endif + index &= ~static_cast(BLOCK_SIZE - 1); + localBlockIndex = blockIndex.load(std::memory_order_acquire); + auto tail = localBlockIndex->tail.load(std::memory_order_acquire); + auto tailBase = localBlockIndex->index[tail]->key.load(std::memory_order_relaxed); + assert(tailBase != INVALID_BLOCK_BASE); + // Note: Must use division instead of shift because the index may wrap around, causing a negative + // offset, whose negativity we want to preserve + auto offset = static_cast(static_cast::type>(index - tailBase) / BLOCK_SIZE); + size_t idx = (tail + offset) & (localBlockIndex->capacity - 1); + assert(localBlockIndex->index[idx]->key.load(std::memory_order_relaxed) == index && localBlockIndex->index[idx]->value.load(std::memory_order_relaxed) != nullptr); + return idx; + } + + bool new_block_index() + { + auto prev = blockIndex.load(std::memory_order_relaxed); + size_t prevCapacity = prev == nullptr ? 0 : prev->capacity; + auto entryCount = prev == nullptr ? nextBlockIndexCapacity : prevCapacity; + auto raw = static_cast((Traits::malloc)( + sizeof(BlockIndexHeader) + + std::alignment_of::value - 1 + sizeof(BlockIndexEntry) * entryCount + + std::alignment_of::value - 1 + sizeof(BlockIndexEntry*) * nextBlockIndexCapacity)); + if (raw == nullptr) { + return false; + } + + auto header = new (raw) BlockIndexHeader; + auto entries = reinterpret_cast(details::align_for(raw + sizeof(BlockIndexHeader))); + auto index = reinterpret_cast(details::align_for(reinterpret_cast(entries) + sizeof(BlockIndexEntry) * entryCount)); + if (prev != nullptr) { + auto prevTail = prev->tail.load(std::memory_order_relaxed); + auto prevPos = prevTail; + size_t i = 0; + do { + prevPos = (prevPos + 1) & (prev->capacity - 1); + index[i++] = prev->index[prevPos]; + } while (prevPos != prevTail); + assert(i == prevCapacity); + } + for (size_t i = 0; i != entryCount; ++i) { + new (entries + i) BlockIndexEntry; + entries[i].key.store(INVALID_BLOCK_BASE, std::memory_order_relaxed); + index[prevCapacity + i] = entries + i; + } + header->prev = prev; + header->entries = entries; + header->index = index; + header->capacity = nextBlockIndexCapacity; + header->tail.store((prevCapacity - 1) & (nextBlockIndexCapacity - 1), std::memory_order_relaxed); + + blockIndex.store(header, std::memory_order_release); + + nextBlockIndexCapacity <<= 1; + + return true; + } + + private: + size_t nextBlockIndexCapacity; + std::atomic blockIndex; + +#ifdef MOODYCAMEL_CPP11_THREAD_LOCAL_SUPPORTED + public: + details::ThreadExitListener threadExitListener; + private: +#endif + +#ifdef MOODYCAMEL_QUEUE_INTERNAL_DEBUG + public: + ImplicitProducer* nextImplicitProducer; + private: +#endif + +#ifdef MCDBGQ_NOLOCKFREE_IMPLICITPRODBLOCKINDEX + mutable debug::DebugMutex mutex; +#endif +#ifdef MCDBGQ_TRACKMEM + friend struct MemStats; +#endif + }; + + + ////////////////////////////////// + // Block pool manipulation + ////////////////////////////////// + + void populate_initial_block_list(size_t blockCount) + { + initialBlockPoolSize = blockCount; + if (initialBlockPoolSize == 0) { + initialBlockPool = nullptr; + return; + } + + initialBlockPool = create_array(blockCount); + if (initialBlockPool == nullptr) { + initialBlockPoolSize = 0; + } + for (size_t i = 0; i < initialBlockPoolSize; ++i) { + initialBlockPool[i].dynamicallyAllocated = false; + } + } + + inline Block* try_get_block_from_initial_pool() + { + if (initialBlockPoolIndex.load(std::memory_order_relaxed) >= initialBlockPoolSize) { + return nullptr; + } + + auto index = initialBlockPoolIndex.fetch_add(1, std::memory_order_relaxed); + + return index < initialBlockPoolSize ? (initialBlockPool + index) : nullptr; + } + + inline void add_block_to_free_list(Block* block) + { +#ifdef MCDBGQ_TRACKMEM + block->owner = nullptr; +#endif + freeList.add(block); + } + + inline void add_blocks_to_free_list(Block* block) + { + while (block != nullptr) { + auto next = block->next; + add_block_to_free_list(block); + block = next; + } + } + + inline Block* try_get_block_from_free_list() + { + return freeList.try_get(); + } + + // Gets a free block from one of the memory pools, or allocates a new one (if applicable) + template + Block* requisition_block() + { + auto block = try_get_block_from_initial_pool(); + if (block != nullptr) { + return block; + } + + block = try_get_block_from_free_list(); + if (block != nullptr) { + return block; + } + + if (canAlloc == CanAlloc) { + return create(); + } + + return nullptr; + } + + +#ifdef MCDBGQ_TRACKMEM + public: + struct MemStats { + size_t allocatedBlocks; + size_t usedBlocks; + size_t freeBlocks; + size_t ownedBlocksExplicit; + size_t ownedBlocksImplicit; + size_t implicitProducers; + size_t explicitProducers; + size_t elementsEnqueued; + size_t blockClassBytes; + size_t queueClassBytes; + size_t implicitBlockIndexBytes; + size_t explicitBlockIndexBytes; + + friend class ConcurrentQueue; + + private: + static MemStats getFor(ConcurrentQueue* q) + { + MemStats stats = { 0 }; + + stats.elementsEnqueued = q->size_approx(); + + auto block = q->freeList.head_unsafe(); + while (block != nullptr) { + ++stats.allocatedBlocks; + ++stats.freeBlocks; + block = block->freeListNext.load(std::memory_order_relaxed); + } + + for (auto ptr = q->producerListTail.load(std::memory_order_acquire); ptr != nullptr; ptr = ptr->next_prod()) { + bool implicit = dynamic_cast(ptr) != nullptr; + stats.implicitProducers += implicit ? 1 : 0; + stats.explicitProducers += implicit ? 0 : 1; + + if (implicit) { + auto prod = static_cast(ptr); + stats.queueClassBytes += sizeof(ImplicitProducer); + auto head = prod->headIndex.load(std::memory_order_relaxed); + auto tail = prod->tailIndex.load(std::memory_order_relaxed); + auto hash = prod->blockIndex.load(std::memory_order_relaxed); + if (hash != nullptr) { + for (size_t i = 0; i != hash->capacity; ++i) { + if (hash->index[i]->key.load(std::memory_order_relaxed) != ImplicitProducer::INVALID_BLOCK_BASE && hash->index[i]->value.load(std::memory_order_relaxed) != nullptr) { + ++stats.allocatedBlocks; + ++stats.ownedBlocksImplicit; + } + } + stats.implicitBlockIndexBytes += hash->capacity * sizeof(typename ImplicitProducer::BlockIndexEntry); + for (; hash != nullptr; hash = hash->prev) { + stats.implicitBlockIndexBytes += sizeof(typename ImplicitProducer::BlockIndexHeader) + hash->capacity * sizeof(typename ImplicitProducer::BlockIndexEntry*); + } + } + for (; details::circular_less_than(head, tail); head += BLOCK_SIZE) { + //auto block = prod->get_block_index_entry_for_index(head); + ++stats.usedBlocks; + } + } + else { + auto prod = static_cast(ptr); + stats.queueClassBytes += sizeof(ExplicitProducer); + auto tailBlock = prod->tailBlock; + bool wasNonEmpty = false; + if (tailBlock != nullptr) { + auto block = tailBlock; + do { + ++stats.allocatedBlocks; + if (!block->ConcurrentQueue::Block::template is_empty() || wasNonEmpty) { + ++stats.usedBlocks; + wasNonEmpty = wasNonEmpty || block != tailBlock; + } + ++stats.ownedBlocksExplicit; + block = block->next; + } while (block != tailBlock); + } + auto index = prod->blockIndex.load(std::memory_order_relaxed); + while (index != nullptr) { + stats.explicitBlockIndexBytes += sizeof(typename ExplicitProducer::BlockIndexHeader) + index->size * sizeof(typename ExplicitProducer::BlockIndexEntry); + index = static_cast(index->prev); + } + } + } + + auto freeOnInitialPool = q->initialBlockPoolIndex.load(std::memory_order_relaxed) >= q->initialBlockPoolSize ? 0 : q->initialBlockPoolSize - q->initialBlockPoolIndex.load(std::memory_order_relaxed); + stats.allocatedBlocks += freeOnInitialPool; + stats.freeBlocks += freeOnInitialPool; + + stats.blockClassBytes = sizeof(Block) * stats.allocatedBlocks; + stats.queueClassBytes += sizeof(ConcurrentQueue); + + return stats; + } + }; + + // For debugging only. Not thread-safe. + MemStats getMemStats() + { + return MemStats::getFor(this); + } + private: + friend struct MemStats; +#endif + + + ////////////////////////////////// + // Producer list manipulation + ////////////////////////////////// + + ProducerBase* recycle_or_create_producer(bool isExplicit) + { + bool recycled; + return recycle_or_create_producer(isExplicit, recycled); + } + + ProducerBase* recycle_or_create_producer(bool isExplicit, bool& recycled) + { +#ifdef MCDBGQ_NOLOCKFREE_IMPLICITPRODHASH + debug::DebugLock lock(implicitProdMutex); +#endif + // Try to re-use one first + for (auto ptr = producerListTail.load(std::memory_order_acquire); ptr != nullptr; ptr = ptr->next_prod()) { + if (ptr->inactive.load(std::memory_order_relaxed) && ptr->isExplicit == isExplicit) { + bool expected = true; + if (ptr->inactive.compare_exchange_strong(expected, /* desired */ false, std::memory_order_acquire, std::memory_order_relaxed)) { + // We caught one! It's been marked as activated, the caller can have it + recycled = true; + return ptr; + } + } + } + + recycled = false; + return add_producer(isExplicit ? static_cast(create(this)) : create(this)); + } + + ProducerBase* add_producer(ProducerBase* producer) + { + // Handle failed memory allocation + if (producer == nullptr) { + return nullptr; + } + + producerCount.fetch_add(1, std::memory_order_relaxed); + + // Add it to the lock-free list + auto prevTail = producerListTail.load(std::memory_order_relaxed); + do { + producer->next = prevTail; + } while (!producerListTail.compare_exchange_weak(prevTail, producer, std::memory_order_release, std::memory_order_relaxed)); + +#ifdef MOODYCAMEL_QUEUE_INTERNAL_DEBUG + if (producer->isExplicit) { + auto prevTailExplicit = explicitProducers.load(std::memory_order_relaxed); + do { + static_cast(producer)->nextExplicitProducer = prevTailExplicit; + } while (!explicitProducers.compare_exchange_weak(prevTailExplicit, static_cast(producer), std::memory_order_release, std::memory_order_relaxed)); + } + else { + auto prevTailImplicit = implicitProducers.load(std::memory_order_relaxed); + do { + static_cast(producer)->nextImplicitProducer = prevTailImplicit; + } while (!implicitProducers.compare_exchange_weak(prevTailImplicit, static_cast(producer), std::memory_order_release, std::memory_order_relaxed)); + } +#endif + + return producer; + } + + void reown_producers() + { + // After another instance is moved-into/swapped-with this one, all the + // producers we stole still think their parents are the other queue. + // So fix them up! + for (auto ptr = producerListTail.load(std::memory_order_relaxed); ptr != nullptr; ptr = ptr->next_prod()) { + ptr->parent = this; + } + } + + + ////////////////////////////////// + // Implicit producer hash + ////////////////////////////////// + + struct ImplicitProducerKVP + { + std::atomic key; + ImplicitProducer* value; // No need for atomicity since it's only read by the thread that sets it in the first place + + ImplicitProducerKVP() : value(nullptr) { } + + ImplicitProducerKVP(ImplicitProducerKVP&& other) MOODYCAMEL_NOEXCEPT + { + key.store(other.key.load(std::memory_order_relaxed), std::memory_order_relaxed); + value = other.value; + } + + inline ImplicitProducerKVP& operator=(ImplicitProducerKVP&& other) MOODYCAMEL_NOEXCEPT + { + swap(other); + return *this; + } + + inline void swap(ImplicitProducerKVP& other) MOODYCAMEL_NOEXCEPT + { + if (this != &other) { + details::swap_relaxed(key, other.key); + std::swap(value, other.value); + } + } + }; + + template + friend void moodycamel::swap(typename ConcurrentQueue::ImplicitProducerKVP&, typename ConcurrentQueue::ImplicitProducerKVP&) MOODYCAMEL_NOEXCEPT; + + struct ImplicitProducerHash + { + size_t capacity; + ImplicitProducerKVP* entries; + ImplicitProducerHash* prev; + }; + + inline void populate_initial_implicit_producer_hash() + { + if (INITIAL_IMPLICIT_PRODUCER_HASH_SIZE == 0) return; + + implicitProducerHashCount.store(0, std::memory_order_relaxed); + auto hash = &initialImplicitProducerHash; + hash->capacity = INITIAL_IMPLICIT_PRODUCER_HASH_SIZE; + hash->entries = &initialImplicitProducerHashEntries[0]; + for (size_t i = 0; i != INITIAL_IMPLICIT_PRODUCER_HASH_SIZE; ++i) { + initialImplicitProducerHashEntries[i].key.store(details::invalid_thread_id, std::memory_order_relaxed); + } + hash->prev = nullptr; + implicitProducerHash.store(hash, std::memory_order_relaxed); + } + + void swap_implicit_producer_hashes(ConcurrentQueue& other) + { + if (INITIAL_IMPLICIT_PRODUCER_HASH_SIZE == 0) return; + + // Swap (assumes our implicit producer hash is initialized) + initialImplicitProducerHashEntries.swap(other.initialImplicitProducerHashEntries); + initialImplicitProducerHash.entries = &initialImplicitProducerHashEntries[0]; + other.initialImplicitProducerHash.entries = &other.initialImplicitProducerHashEntries[0]; + + details::swap_relaxed(implicitProducerHashCount, other.implicitProducerHashCount); + + details::swap_relaxed(implicitProducerHash, other.implicitProducerHash); + if (implicitProducerHash.load(std::memory_order_relaxed) == &other.initialImplicitProducerHash) { + implicitProducerHash.store(&initialImplicitProducerHash, std::memory_order_relaxed); + } + else { + ImplicitProducerHash* hash; + for (hash = implicitProducerHash.load(std::memory_order_relaxed); hash->prev != &other.initialImplicitProducerHash; hash = hash->prev) { + continue; + } + hash->prev = &initialImplicitProducerHash; + } + if (other.implicitProducerHash.load(std::memory_order_relaxed) == &initialImplicitProducerHash) { + other.implicitProducerHash.store(&other.initialImplicitProducerHash, std::memory_order_relaxed); + } + else { + ImplicitProducerHash* hash; + for (hash = other.implicitProducerHash.load(std::memory_order_relaxed); hash->prev != &initialImplicitProducerHash; hash = hash->prev) { + continue; + } + hash->prev = &other.initialImplicitProducerHash; + } + } + + // Only fails (returns nullptr) if memory allocation fails + ImplicitProducer* get_or_add_implicit_producer() + { + // Note that since the data is essentially thread-local (key is thread ID), + // there's a reduced need for fences (memory ordering is already consistent + // for any individual thread), except for the current table itself. + + // Start by looking for the thread ID in the current and all previous hash tables. + // If it's not found, it must not be in there yet, since this same thread would + // have added it previously to one of the tables that we traversed. + + // Code and algorithm adapted from http://preshing.com/20130605/the-worlds-simplest-lock-free-hash-table + +#ifdef MCDBGQ_NOLOCKFREE_IMPLICITPRODHASH + debug::DebugLock lock(implicitProdMutex); +#endif + + auto id = details::thread_id(); + auto hashedId = details::hash_thread_id(id); + + auto mainHash = implicitProducerHash.load(std::memory_order_acquire); + for (auto hash = mainHash; hash != nullptr; hash = hash->prev) { + // Look for the id in this hash + auto index = hashedId; + while (true) { // Not an infinite loop because at least one slot is free in the hash table + index &= hash->capacity - 1; + + auto probedKey = hash->entries[index].key.load(std::memory_order_relaxed); + if (probedKey == id) { + // Found it! If we had to search several hashes deep, though, we should lazily add it + // to the current main hash table to avoid the extended search next time. + // Note there's guaranteed to be room in the current hash table since every subsequent + // table implicitly reserves space for all previous tables (there's only one + // implicitProducerHashCount). + auto value = hash->entries[index].value; + if (hash != mainHash) { + index = hashedId; + while (true) { + index &= mainHash->capacity - 1; + probedKey = mainHash->entries[index].key.load(std::memory_order_relaxed); + auto empty = details::invalid_thread_id; +#ifdef MOODYCAMEL_CPP11_THREAD_LOCAL_SUPPORTED + auto reusable = details::invalid_thread_id2; + if ((probedKey == empty && mainHash->entries[index].key.compare_exchange_strong(empty, id, std::memory_order_relaxed, std::memory_order_relaxed)) || + (probedKey == reusable && mainHash->entries[index].key.compare_exchange_strong(reusable, id, std::memory_order_acquire, std::memory_order_acquire))) { +#else + if ((probedKey == empty && mainHash->entries[index].key.compare_exchange_strong(empty, id, std::memory_order_relaxed, std::memory_order_relaxed))) { +#endif + mainHash->entries[index].value = value; + break; + } + ++index; + } + } + + return value; + } + if (probedKey == details::invalid_thread_id) { + break; // Not in this hash table + } + ++index; + } + } + + // Insert! + auto newCount = 1 + implicitProducerHashCount.fetch_add(1, std::memory_order_relaxed); + while (true) { + // NOLINTNEXTLINE(clang-analyzer-core.NullDereference) + if (newCount >= (mainHash->capacity >> 1) && !implicitProducerHashResizeInProgress.test_and_set(std::memory_order_acquire)) { + // We've acquired the resize lock, try to allocate a bigger hash table. + // Note the acquire fence synchronizes with the release fence at the end of this block, and hence when + // we reload implicitProducerHash it must be the most recent version (it only gets changed within this + // locked block). + mainHash = implicitProducerHash.load(std::memory_order_acquire); + if (newCount >= (mainHash->capacity >> 1)) { + auto newCapacity = mainHash->capacity << 1; + while (newCount >= (newCapacity >> 1)) { + newCapacity <<= 1; + } + auto raw = static_cast((Traits::malloc)(sizeof(ImplicitProducerHash) + std::alignment_of::value - 1 + sizeof(ImplicitProducerKVP) * newCapacity)); + if (raw == nullptr) { + // Allocation failed + implicitProducerHashCount.fetch_sub(1, std::memory_order_relaxed); + implicitProducerHashResizeInProgress.clear(std::memory_order_relaxed); + return nullptr; + } + + auto newHash = new (raw) ImplicitProducerHash; + newHash->capacity = newCapacity; + newHash->entries = reinterpret_cast(details::align_for(raw + sizeof(ImplicitProducerHash))); + for (size_t i = 0; i != newCapacity; ++i) { + new (newHash->entries + i) ImplicitProducerKVP; + newHash->entries[i].key.store(details::invalid_thread_id, std::memory_order_relaxed); + } + newHash->prev = mainHash; + implicitProducerHash.store(newHash, std::memory_order_release); + implicitProducerHashResizeInProgress.clear(std::memory_order_release); + mainHash = newHash; + } + else { + implicitProducerHashResizeInProgress.clear(std::memory_order_release); + } + } + + // If it's < three-quarters full, add to the old one anyway so that we don't have to wait for the next table + // to finish being allocated by another thread (and if we just finished allocating above, the condition will + // always be true) + if (newCount < (mainHash->capacity >> 1) + (mainHash->capacity >> 2)) { + bool recycled; + auto producer = static_cast(recycle_or_create_producer(false, recycled)); + if (producer == nullptr) { + implicitProducerHashCount.fetch_sub(1, std::memory_order_relaxed); + return nullptr; + } + if (recycled) { + implicitProducerHashCount.fetch_sub(1, std::memory_order_relaxed); + } + +#ifdef MOODYCAMEL_CPP11_THREAD_LOCAL_SUPPORTED + producer->threadExitListener.callback = &ConcurrentQueue::implicit_producer_thread_exited_callback; + producer->threadExitListener.userData = producer; + details::ThreadExitNotifier::subscribe(&producer->threadExitListener); +#endif + + auto index = hashedId; + while (true) { + index &= mainHash->capacity - 1; + auto probedKey = mainHash->entries[index].key.load(std::memory_order_relaxed); + + auto empty = details::invalid_thread_id; +#ifdef MOODYCAMEL_CPP11_THREAD_LOCAL_SUPPORTED + auto reusable = details::invalid_thread_id2; + if ((probedKey == empty && mainHash->entries[index].key.compare_exchange_strong(empty, id, std::memory_order_relaxed, std::memory_order_relaxed)) || + (probedKey == reusable && mainHash->entries[index].key.compare_exchange_strong(reusable, id, std::memory_order_acquire, std::memory_order_acquire))) { +#else + if ((probedKey == empty && mainHash->entries[index].key.compare_exchange_strong(empty, id, std::memory_order_relaxed, std::memory_order_relaxed))) { +#endif + mainHash->entries[index].value = producer; + break; + } + ++index; + } + return producer; + } + + // Hmm, the old hash is quite full and somebody else is busy allocating a new one. + // We need to wait for the allocating thread to finish (if it succeeds, we add, if not, + // we try to allocate ourselves). + mainHash = implicitProducerHash.load(std::memory_order_acquire); + } + } + +#ifdef MOODYCAMEL_CPP11_THREAD_LOCAL_SUPPORTED + void implicit_producer_thread_exited(ImplicitProducer* producer) + { + // Remove from thread exit listeners + details::ThreadExitNotifier::unsubscribe(&producer->threadExitListener); + + // Remove from hash +#ifdef MCDBGQ_NOLOCKFREE_IMPLICITPRODHASH + debug::DebugLock lock(implicitProdMutex); +#endif + auto hash = implicitProducerHash.load(std::memory_order_acquire); + assert(hash != nullptr); // The thread exit listener is only registered if we were added to a hash in the first place + auto id = details::thread_id(); + auto hashedId = details::hash_thread_id(id); + details::thread_id_t probedKey; + + // We need to traverse all the hashes just in case other threads aren't on the current one yet and are + // trying to add an entry thinking there's a free slot (because they reused a producer) + for (; hash != nullptr; hash = hash->prev) { + auto index = hashedId; + do { + index &= hash->capacity - 1; + probedKey = hash->entries[index].key.load(std::memory_order_relaxed); + if (probedKey == id) { + hash->entries[index].key.store(details::invalid_thread_id2, std::memory_order_release); + break; + } + ++index; + } while (probedKey != details::invalid_thread_id); // Can happen if the hash has changed but we weren't put back in it yet, or if we weren't added to this hash in the first place + } + + // Mark the queue as being recyclable + producer->inactive.store(true, std::memory_order_release); + } + + static void implicit_producer_thread_exited_callback(void* userData) + { + auto producer = static_cast(userData); + auto queue = producer->parent; + queue->implicit_producer_thread_exited(producer); + } +#endif + + ////////////////////////////////// + // Utility functions + ////////////////////////////////// + + template + static inline U* create_array(size_t count) + { + assert(count > 0); + auto p = static_cast((Traits::malloc)(sizeof(U) * count)); + if (p == nullptr) { + return nullptr; + } + + for (size_t i = 0; i != count; ++i) { + new (p + i) U(); + } + return p; + } + + template + static inline void destroy_array(U* p, size_t count) + { + if (p != nullptr) { + assert(count > 0); + for (size_t i = count; i != 0; ) { + (p + --i)->~U(); + } + (Traits::free)(p); + } + } + + template + static inline U* create() + { + auto p = (Traits::malloc)(sizeof(U)); + return p != nullptr ? new (p) U : nullptr; + } + + template + static inline U* create(A1&& a1) + { + auto p = (Traits::malloc)(sizeof(U)); + return p != nullptr ? new (p) U(std::forward(a1)) : nullptr; + } + + template + static inline void destroy(U* p) + { + if (p != nullptr) { + p->~U(); + } + (Traits::free)(p); + } + +private: + std::atomic producerListTail; + std::atomic producerCount; + + std::atomic initialBlockPoolIndex; + Block* initialBlockPool; + size_t initialBlockPoolSize; + +#if !MCDBGQ_USEDEBUGFREELIST + FreeList freeList; +#else + debug::DebugFreeList freeList; +#endif + + std::atomic implicitProducerHash; + std::atomic implicitProducerHashCount; // Number of slots logically used + ImplicitProducerHash initialImplicitProducerHash; + std::array initialImplicitProducerHashEntries; + std::atomic_flag implicitProducerHashResizeInProgress; + + std::atomic nextExplicitConsumerId; + std::atomic globalExplicitConsumerOffset; + +#ifdef MCDBGQ_NOLOCKFREE_IMPLICITPRODHASH + debug::DebugMutex implicitProdMutex; +#endif + +#ifdef MOODYCAMEL_QUEUE_INTERNAL_DEBUG + std::atomic explicitProducers; + std::atomic implicitProducers; +#endif +}; + + +template +ProducerToken::ProducerToken(ConcurrentQueue& queue) + : producer(queue.recycle_or_create_producer(true)) +{ + if (producer != nullptr) { + producer->token = this; + } +} + +template +ProducerToken::ProducerToken(BlockingConcurrentQueue& queue) + : producer(reinterpret_cast*>(&queue)->recycle_or_create_producer(true)) +{ + if (producer != nullptr) { + producer->token = this; + } +} + +template +ConsumerToken::ConsumerToken(ConcurrentQueue& queue) + : itemsConsumedFromCurrent(0), currentProducer(nullptr), desiredProducer(nullptr) +{ + initialOffset = queue.nextExplicitConsumerId.fetch_add(1, std::memory_order_release); + lastKnownGlobalOffset = -1; +} + +template +ConsumerToken::ConsumerToken(BlockingConcurrentQueue& queue) + : itemsConsumedFromCurrent(0), currentProducer(nullptr), desiredProducer(nullptr) +{ + initialOffset = reinterpret_cast*>(&queue)->nextExplicitConsumerId.fetch_add(1, std::memory_order_release); + lastKnownGlobalOffset = -1; +} + +template +inline void swap(ConcurrentQueue& a, ConcurrentQueue& b) MOODYCAMEL_NOEXCEPT +{ + a.swap(b); +} + +inline void swap(ProducerToken& a, ProducerToken& b) MOODYCAMEL_NOEXCEPT +{ + a.swap(b); +} + +inline void swap(ConsumerToken& a, ConsumerToken& b) MOODYCAMEL_NOEXCEPT +{ + a.swap(b); +} + +template +inline void swap(typename ConcurrentQueue::ImplicitProducerKVP& a, typename ConcurrentQueue::ImplicitProducerKVP& b) MOODYCAMEL_NOEXCEPT +{ + a.swap(b); +} + +} + +#if defined(__GNUC__) +#pragma GCC diagnostic pop +#endif diff --git a/src/ctx.cpp b/src/ctx.cpp index fdf0c163ac..edcc1aae31 100644 --- a/src/ctx.cpp +++ b/src/ctx.cpp @@ -47,6 +47,7 @@ #include "err.hpp" #include "msg.hpp" #include "random.hpp" +#include "allocator.hpp" #ifdef ZMQ_HAVE_VMCI #include @@ -279,6 +280,18 @@ int zmq::ctx_t::set (int option_, const void *optval_, size_t optvallen_) } break; + /*case ZMQ_MSG_ALLOCATOR: { + if (optvallen_ == sizeof (zmq::allocator_t)) { + const zmq::allocator_t *all = + static_cast (optval_); + if (all->check_tag ()) { + _allocator = const_cast (all); + return 0; + } + } + break; + }*/ + default: { return thread_ctx_t::set (option_, optval_, optvallen_); } @@ -349,6 +362,9 @@ int zmq::ctx_t::get (int option_, void *optval_, size_t *optvallen_) return 0; } break; + /* + case ZMQ_MSG_ALLOCATOR: { + } break;*/ default: { return thread_ctx_t::get (option_, optval_, optvallen_); diff --git a/src/ctx.hpp b/src/ctx.hpp index 9aef843485..e8975d16eb 100644 --- a/src/ctx.hpp +++ b/src/ctx.hpp @@ -35,6 +35,7 @@ #include #include +//#include "allocator.hpp" #include "mailbox.hpp" #include "array.hpp" #include "config.hpp" @@ -220,6 +221,9 @@ class ctx_t : public thread_ctx_t // Synchronisation of access to the list of inproc endpoints. mutex_t _endpoints_sync; + // Allocator for messages + //allocator_t *_allocator; + // Maximum socket ID. static atomic_counter_t max_socket_id; diff --git a/src/msg.cpp b/src/msg.cpp index dc1081c4c2..0a6b7a4570 100644 --- a/src/msg.cpp +++ b/src/msg.cpp @@ -39,6 +39,7 @@ #include "likely.hpp" #include "metadata.hpp" #include "err.hpp" +#include "allocator.hpp" // Check whether the sizes of public representation of the message (zmq_msg_t) // and private representation of the message (zmq::msg_t) match. @@ -47,8 +48,6 @@ typedef char zmq_msg_size_check[2 * ((sizeof (zmq::msg_t) == sizeof (zmq_msg_t)) != 0) - 1]; -#define ALLOCATE_CONTENT_T_INSIDE_PROVIDED_BUFFER (1) - bool zmq::msg_t::check () const { @@ -100,6 +99,7 @@ int zmq::msg_t::init_size (size_t size_) _u.lmsg.metadata = NULL; _u.lmsg.type = type_lmsg; _u.lmsg.flags = 0; + _u.lmsg.allocator_was_used = 0; _u.lmsg.group[0] = '\0'; _u.lmsg.routing_id = 0; _u.lmsg.content = NULL; @@ -167,28 +167,18 @@ int zmq::msg_t::init_data (void *data_, _u.lmsg.metadata = NULL; _u.lmsg.type = type_lmsg; _u.lmsg.flags = 0; + _u.lmsg.allocator_was_used = 0; _u.lmsg.group[0] = '\0'; _u.lmsg.routing_id = 0; -#if ALLOCATE_CONTENT_T_INSIDE_PROVIDED_BUFFER - zmq_assert (size_ > sizeof (content_t)); - _u.lmsg.content = reinterpret_cast (data_); -#else _u.lmsg.content = static_cast (malloc (sizeof (content_t))); -#endif if (!_u.lmsg.content) { errno = ENOMEM; return -1; } -#if ALLOCATE_CONTENT_T_INSIDE_PROVIDED_BUFFER - uint8_t *data_bytes = (uint8_t *) data_; - _u.lmsg.content->data = data_bytes + sizeof (content_t); - _u.lmsg.content->size = size_ - sizeof (content_t); -#else _u.lmsg.content->data = data_; _u.lmsg.content->size = size_; -#endif _u.lmsg.content->ffn = ffn_; _u.lmsg.content->hint = hint_; new (&_u.lmsg.content->refcnt) zmq::atomic_counter_t (); @@ -196,6 +186,33 @@ int zmq::msg_t::init_data (void *data_, return 0; } +int zmq::msg_t::init_from_allocator (size_t size_, zmq::allocator_t *alloc_) +{ + zmq_assert (alloc_ != NULL && size_ != 0); + + _u.lmsg.metadata = NULL; + _u.lmsg.type = type_lmsg; + _u.lmsg.flags = 0; + _u.lmsg.allocator_was_used = 1; + _u.lmsg.group[0] = '\0'; + _u.lmsg.routing_id = 0; + _u.lmsg.content = reinterpret_cast ( + alloc_->allocate (size_ + sizeof (content_t))); + + if (!_u.lmsg.content) { + errno = ENOMEM; + return -1; + } + + _u.lmsg.content->data = _u.lmsg.content + 1; + _u.lmsg.content->size = size_; + _u.lmsg.content->ffn = (zmq_free_fn *) alloc_->deallocate_msg; + _u.lmsg.content->hint = alloc_; + new (&_u.lmsg.content->refcnt) zmq::atomic_counter_t (); + + return 0; +} + int zmq::msg_t::init_delimiter () { _u.delimiter.metadata = NULL; @@ -242,23 +259,25 @@ int zmq::msg_t::close () // We used "placement new" operator to initialize the reference // counter so we call the destructor explicitly now. _u.lmsg.content->refcnt.~atomic_counter_t (); -#if ALLOCATE_CONTENT_T_INSIDE_PROVIDED_BUFFER - // take a local copy since we are going to remove (through the user-provided deallocator) - // the whole malloc'ed buffer, including the content_t block itself! - // NOTE: this copy should not be strictly needed but it's here just to help debugging: - content_t content; - content.data = _u.lmsg.content->data; - content.size = _u.lmsg.content->size; - content.ffn = _u.lmsg.content->ffn; - content.hint = _u.lmsg.content->hint; - if (content.ffn) - content.ffn (content.data, content.hint); -#else - if (_u.lmsg.content->ffn) - _u.lmsg.content->ffn (_u.lmsg.content->data, - _u.lmsg.content->hint); - free (_u.lmsg.content); -#endif + + if (_u.lmsg.allocator_was_used) { + // take a local copy since we are going to remove (through the user-provided deallocator) + // the whole malloc'ed buffer, including the content_t block itself! + // NOTE: this copy should not be strictly needed but it's here just to help debugging: + content_t content; + content.data = _u.lmsg.content->data; + content.size = _u.lmsg.content->size; + content.ffn = _u.lmsg.content->ffn; + content.hint = _u.lmsg.content->hint; + if (content.ffn) + /* return to the allocator the memory starting from the content_t struct */ + content.ffn (_u.lmsg.content, content.hint); + } else { + if (_u.lmsg.content->ffn) + _u.lmsg.content->ffn (_u.lmsg.content->data, + _u.lmsg.content->hint); + free (_u.lmsg.content); + } } } diff --git a/src/msg.hpp b/src/msg.hpp index c4407c286c..bcbac866d4 100644 --- a/src/msg.hpp +++ b/src/msg.hpp @@ -51,6 +51,8 @@ typedef void(msg_free_fn) (void *data_, void *hint_); namespace zmq { +class allocator_t; + // Note that this structure needs to be explicitly constructed // (init functions) and destructed (close function). @@ -105,6 +107,7 @@ class msg_t size_t size_, msg_free_fn *ffn_, void *hint_); + int init_from_allocator (size_t size_, zmq::allocator_t *alloc_); int init_delimiter (); int init_join (); int init_leave (); @@ -236,9 +239,10 @@ class msg_t { metadata_t *metadata; content_t *content; + unsigned char allocator_was_used; // boolean flag unsigned char unused[msg_t_size - (sizeof (metadata_t *) + sizeof (content_t *) - + 2 + 16 + sizeof (uint32_t))]; + + 3 + 16 + sizeof (uint32_t))]; unsigned char type; unsigned char flags; char group[16]; diff --git a/src/zmq.cpp b/src/zmq.cpp index 0931e61f62..c8b1dc4041 100644 --- a/src/zmq.cpp +++ b/src/zmq.cpp @@ -95,6 +95,7 @@ struct iovec #include "timers.hpp" #include "ip.hpp" #include "address.hpp" +#include "allocator.hpp" #if defined ZMQ_HAVE_OPENPGM #define __PGM_WININT_H__ @@ -215,6 +216,36 @@ int zmq_ctx_get_ext (void *ctx_, int option_, void *optval_, size_t *optvallen_) } +// New allocator API + +void *zmq_msg_allocator_new (int type_) +{ + zmq::allocator_t *pool = new (std::nothrow) zmq::allocator_t; + if (!pool) { + errno = ENOMEM; + return NULL; + } + + pool->init (type_); + return pool; +} + +int zmq_msg_allocator_destroy (void **allocator_) +{ + if (allocator_) { + zmq::allocator_t *const allocator = + static_cast (*allocator_); + if (allocator && allocator->check_tag ()) { + delete allocator; + *allocator_ = NULL; + return 0; + } + } + errno = EFAULT; + return -1; +} + + // Stable/legacy context API void *zmq_init (int io_threads_) @@ -600,6 +631,13 @@ int zmq_msg_init_size (zmq_msg_t *msg_, size_t size_) return (reinterpret_cast (msg_))->init_size (size_); } +int zmq_msg_init_allocator (zmq_msg_t *msg_, size_t size_, void *allocator_) +{ + return (reinterpret_cast (msg_)) + ->init_from_allocator (size_, + reinterpret_cast (allocator_)); +} + int zmq_msg_init_data ( zmq_msg_t *msg_, void *data_, size_t size_, zmq_free_fn *ffn_, void *hint_) { From 18c52c4648116590003ba0aa5f1bcf3c9db7bee5 Mon Sep 17 00:00:00 2001 From: Francesco Montorsi Date: Thu, 29 Aug 2019 00:42:02 +0200 Subject: [PATCH 07/11] Remove changes related to graph generation --- perf/generate_csv.sh | 33 +-------------------------- perf/generate_graphs.py | 49 +++++++++++++---------------------------- 2 files changed, 16 insertions(+), 66 deletions(-) diff --git a/perf/generate_csv.sh b/perf/generate_csv.sh index da8ff0a4cd..d307f29e49 100755 --- a/perf/generate_csv.sh +++ b/perf/generate_csv.sh @@ -10,7 +10,6 @@ # export LOCAL_TEST_ENDPOINT="tcp://192.168.1.1:1234" # export REMOTE_TEST_ENDPOINT="tcp://192.168.1.2:1234" # export REMOTE_LIBZMQ_PATH="/home/fmontorsi/libzmq/perf" -# export MESSAGE_SIZE_LIST="8 16 32 64 128 210" # ./generate_csv.sh # @@ -23,7 +22,7 @@ LOCAL_TEST_ENDPOINT=${LOCAL_TEST_ENDPOINT:-tcp://192.168.1.1:1234} REMOTE_TEST_ENDPOINT=${REMOTE_TEST_ENDPOINT:-tcp://192.168.1.2:1234} # constant values: -MESSAGE_SIZE_LIST="${MESSAGE_SIZE_LIST:-8 16 32 64 128 256 512 1024 2048 4096 8192 16384 32768 65536 131072}" +MESSAGE_SIZE_LIST="8 16 32 64 128 256 512 1024 2048 4096 8192 16384 32768 65536 131072" OUTPUT_DIR="results" OUTPUT_FILE_PREFIX="results.txt" OUTPUT_FILE_CSV_PREFIX="results.csv" @@ -48,35 +47,6 @@ function verify_ssh() echo "SSH connection to the remote $REMOTE_IP_SSH is working fine." } -function set_reproducible_tcp_kernel_buff_size() -{ - sysctl -w net.core.rmem_max=8388608 && \ - sysctl -w net.core.wmem_max=8388608 && \ - sysctl -w net.core.rmem_default=65536 && \ - sysctl -w net.core.wmem_default=65536 && \ - sysctl -w net.ipv4.tcp_rmem='4096 87380 8388608' && \ - sysctl -w net.ipv4.tcp_wmem='4096 65536 8388608' && \ - sysctl -w net.ipv4.tcp_mem='8388608 8388608 8388608' && \ - sysctl -w net.ipv4.route.flush=1 - if [ $? -ne 0 ]; then - echo "Failed setting kernel socket buffer sizes LOCALLY" - exit 2 - fi - - ssh $REMOTE_IP_SSH "sysctl -w net.core.rmem_max=8388608 && \ - sysctl -w net.core.wmem_max=8388608 && \ - sysctl -w net.core.rmem_default=65536 && \ - sysctl -w net.core.wmem_default=65536 && \ - sysctl -w net.ipv4.tcp_rmem='4096 87380 8388608' && \ - sysctl -w net.ipv4.tcp_wmem='4096 65536 8388608' && \ - sysctl -w net.ipv4.tcp_mem='8388608 8388608 8388608' && \ - sysctl -w net.ipv4.route.flush=1" - if [ $? -ne 0 ]; then - echo "Failed setting kernel socket buffer sizes on the REMOTE system $REMOTE_IP_SSH" - exit 2 - fi -} - function run_remote_perf_util() { local MESSAGE_SIZE_BYTES="$1" @@ -141,7 +111,6 @@ function generate_output_file() # main: verify_ssh -set_reproducible_tcp_kernel_buff_size THROUGHPUT_CSV_HEADER_LINE="# message_size,message_count,PPS[msg/s],throughput[Mb/s]" diff --git a/perf/generate_graphs.py b/perf/generate_graphs.py index c323e4cce6..20651b7160 100755 --- a/perf/generate_graphs.py +++ b/perf/generate_graphs.py @@ -1,14 +1,20 @@ #!/usr/bin/python3 # -# This script assumes that the set of CSV files produced by "generate_csv.sh" is provided as input. -# -# Usage example: -# export RESULT_DIRECTORY="./results" -# export TCP_LINK_SPEED_GBPS="10" # or 1 or 100 as you like -# ./generate_graphs.py +# This script assumes that the set of CSV files produced by "generate_csv.sh" is provided as input +# and that locally there is the "results" folder. # +# results for TCP: +INPUT_FILE_PUSHPULL_TCP_THROUGHPUT="results/pushpull_tcp_thr_results.csv" +INPUT_FILE_REQREP_TCP_LATENCY="results/reqrep_tcp_lat_results.csv" +TCP_LINK_GPBS=100 + +# results for INPROC: +INPUT_FILE_PUSHPULL_INPROC_THROUGHPUT="results/pushpull_inproc_thr_results.csv" +INPUT_FILE_PUBSUBPROXY_INPROC_THROUGHPUT="results/pubsubproxy_inproc_thr_results.csv" + + # dependencies # # pip3 install matplotlib @@ -16,15 +22,13 @@ import matplotlib.pyplot as plt import numpy as np -import os # functions -def plot_throughput(csv_filename, title, is_tcp=False, tcp_link_speed_gbps=10): +def plot_throughput(csv_filename, title, is_tcp=False): message_size_bytes, message_count, pps, mbps = np.loadtxt(csv_filename, delimiter=',', unpack=True) - print("Generating PNG image file [%s] from CSV results '%s'" % (title, csv_filename)) fig, ax1 = plt.subplots() # PPS axis @@ -40,7 +44,7 @@ def plot_throughput(csv_filename, title, is_tcp=False, tcp_link_speed_gbps=10): ax2.set_ylabel('Throughput [Gb/s]', color=color) ax2.semilogx(message_size_bytes, mbps / 1e3, label='Throughput [Gb/s]', marker='o') if is_tcp: - ax2.set_yticks(np.arange(0, tcp_link_speed_gbps + 1, tcp_link_speed_gbps/10)) + ax2.set_yticks(np.arange(0, TCP_LINK_GPBS + 1, TCP_LINK_GPBS/10)) ax2.tick_params(axis='y', labelcolor=color) ax2.grid(True) @@ -51,8 +55,6 @@ def plot_throughput(csv_filename, title, is_tcp=False, tcp_link_speed_gbps=10): def plot_latency(csv_filename, title): message_size_bytes, message_count, lat = np.loadtxt(csv_filename, delimiter=',', unpack=True) - - print("Generating PNG image file [%s] from CSV results '%s'" % (title, csv_filename)) plt.semilogx(message_size_bytes, lat, label='Latency [us]', marker='o') plt.xlabel('Message size [B]') @@ -65,28 +67,7 @@ def plot_latency(csv_filename, title): # main -try: - result_dir = os.environ['RESULT_DIRECTORY'] -except: - result_dir = "results" # default value - -try: - tcp_link_speed_gbps = int(os.environ['TCP_LINK_SPEED_GBPS']) -except: - tcp_link_speed_gbps = 10 # default value - - - -# result files for TCP: -INPUT_FILE_PUSHPULL_TCP_THROUGHPUT = result_dir + "/pushpull_tcp_thr_results.csv" -INPUT_FILE_REQREP_TCP_LATENCY = result_dir + "/reqrep_tcp_lat_results.csv" - -# results for INPROC: -INPUT_FILE_PUSHPULL_INPROC_THROUGHPUT = result_dir + "/pushpull_inproc_thr_results.csv" -INPUT_FILE_PUBSUBPROXY_INPROC_THROUGHPUT = result_dir + "/pubsubproxy_inproc_thr_results.csv" - -# generate plots -plot_throughput(INPUT_FILE_PUSHPULL_TCP_THROUGHPUT, 'ZeroMQ PUSH/PULL socket throughput, TCP transport', is_tcp=True, tcp_link_speed_gbps=tcp_link_speed_gbps) +plot_throughput(INPUT_FILE_PUSHPULL_TCP_THROUGHPUT, 'ZeroMQ PUSH/PULL socket throughput, TCP transport', is_tcp=True) plot_throughput(INPUT_FILE_PUSHPULL_INPROC_THROUGHPUT, 'ZeroMQ PUSH/PULL socket throughput, INPROC transport') plot_throughput(INPUT_FILE_PUBSUBPROXY_INPROC_THROUGHPUT, 'ZeroMQ PUB/SUB PROXY socket throughput, INPROC transport') plot_latency(INPUT_FILE_REQREP_TCP_LATENCY, 'ZeroMQ REQ/REP socket latency, TCP transport') From a720a311d6712248ba556a4b33f84c5761365982 Mon Sep 17 00:00:00 2001 From: Francesco Montorsi Date: Thu, 29 Aug 2019 00:49:47 +0200 Subject: [PATCH 08/11] allow testing up to 8k msg sizes --- src/allocator.hpp | 13 +++++++++++-- 1 file changed, 11 insertions(+), 2 deletions(-) diff --git a/src/allocator.hpp b/src/allocator.hpp index 8cac7e8584..f9e7001786 100644 --- a/src/allocator.hpp +++ b/src/allocator.hpp @@ -35,7 +35,7 @@ #include "concurrentqueue.h" // FIXME: we need to grow dynamically the mempool -#define MAX_ACTIVE_MESSAGES (8192) +#define MAX_ACTIVE_MESSAGES (16384) namespace zmq { @@ -85,7 +85,16 @@ class global_memory_pool_t return MsgBlock_SizeClass_256; else if (n < 512) return MsgBlock_SizeClass_512; - + else if (n < 1024) + return MsgBlock_SizeClass_1024; + else if (n < 2048) + return MsgBlock_SizeClass_2048; + else if (n < 4096) + return MsgBlock_SizeClass_4096; + else if (n < 8192) + return MsgBlock_SizeClass_8192; + + // size too big return MsgBlock_NumSizeClasses; } From b9e1f016e42ca67413c03699e24f025b6d981f5f Mon Sep 17 00:00:00 2001 From: Francesco Montorsi Date: Fri, 30 Aug 2019 23:51:42 +0200 Subject: [PATCH 09/11] correctly deallocate memory pool blocks --- src/allocator.hpp | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-) diff --git a/src/allocator.hpp b/src/allocator.hpp index f9e7001786..e81c4db17b 100644 --- a/src/allocator.hpp +++ b/src/allocator.hpp @@ -116,7 +116,14 @@ class global_memory_pool_t } } } - ~global_memory_pool_t () {} + ~global_memory_pool_t () + { + // deallocate all message classes + for (int i = 0; i < MsgBlock_NumSizeClasses; i++) { + free (m_storage[i].raw_data); + m_storage[i].raw_data = NULL; + } + } void *allocate_msg (size_t len) // consumer thread: user app thread { From 1649701137fef97ddcd7e76f61635fdd9a23d4dc Mon Sep 17 00:00:00 2001 From: Francesco Montorsi Date: Sat, 31 Aug 2019 00:13:16 +0200 Subject: [PATCH 10/11] fix build with no draft API --- src/allocator.hpp | 4 ++-- src/zmq_draft.h | 11 +++++++++++ 2 files changed, 13 insertions(+), 2 deletions(-) diff --git a/src/allocator.hpp b/src/allocator.hpp index e81c4db17b..3e1352e9e9 100644 --- a/src/allocator.hpp +++ b/src/allocator.hpp @@ -177,10 +177,10 @@ class allocator_t void init (int type_) { _type = type_; } - // allocate() gets called by the consumer thread: the user app thread + // allocate() typically gets called by the consumer thread: the user app thread(s) void *allocate (size_t len); - // deallocate_msg() gets called by the producer thread: the ZMQ background IO thread + // deallocate_msg() typically gets called by the producer thread: the ZMQ background IO thread(s) static void deallocate_msg (void *data_, void *hint_); size_t size () const; diff --git a/src/zmq_draft.h b/src/zmq_draft.h index e558958c5d..46909dc383 100644 --- a/src/zmq_draft.h +++ b/src/zmq_draft.h @@ -71,6 +71,17 @@ int zmq_ctx_get_ext (void *context_, void *optval_, size_t *optvallen_); +/* ZMQ-provided message-pool implementations. */ +// default allocator using malloc/free +#define ZMQ_MSG_ALLOCATOR_DEFAULT 0 +// using internally a SPSC queue (cannot be used with inproc maybe?) or perhaps an MPMC queue anyway +#define ZMQ_MSG_ALLOCATOR_PER_THREAD_POOL 1 +// using internally a MPMC queue +#define ZMQ_MSG_ALLOCATOR_GLOBAL_POOL 2 + +void *zmq_msg_allocator_new (int type_); +int zmq_msg_allocator_destroy (void **allocator_); + /* DRAFT Socket methods. */ int zmq_join (void *s_, const char *group_); int zmq_leave (void *s_, const char *group_); From 0baafa49fb80b850a172efb52b95693ae69f80fb Mon Sep 17 00:00:00 2001 From: Francesco Montorsi Date: Sat, 31 Aug 2019 17:00:33 +0200 Subject: [PATCH 11/11] never use allocator for VSM --- src/msg.cpp | 47 +++++++++++++++++++++++++++++------------------ 1 file changed, 29 insertions(+), 18 deletions(-) diff --git a/src/msg.cpp b/src/msg.cpp index 0a6b7a4570..867de62966 100644 --- a/src/msg.cpp +++ b/src/msg.cpp @@ -190,25 +190,36 @@ int zmq::msg_t::init_from_allocator (size_t size_, zmq::allocator_t *alloc_) { zmq_assert (alloc_ != NULL && size_ != 0); - _u.lmsg.metadata = NULL; - _u.lmsg.type = type_lmsg; - _u.lmsg.flags = 0; - _u.lmsg.allocator_was_used = 1; - _u.lmsg.group[0] = '\0'; - _u.lmsg.routing_id = 0; - _u.lmsg.content = reinterpret_cast ( - alloc_->allocate (size_ + sizeof (content_t))); - - if (!_u.lmsg.content) { - errno = ENOMEM; - return -1; - } + if (size_ <= max_vsm_size) { + // in case we can fit the message data inside the msg_t itself, this option will always + // be fastest rather than using the allocator! + _u.vsm.metadata = NULL; + _u.vsm.type = type_vsm; + _u.vsm.flags = 0; + _u.vsm.size = static_cast (size_); + _u.vsm.group[0] = '\0'; + _u.vsm.routing_id = 0; + } else { + _u.lmsg.metadata = NULL; + _u.lmsg.type = type_lmsg; + _u.lmsg.flags = 0; + _u.lmsg.allocator_was_used = 1; + _u.lmsg.group[0] = '\0'; + _u.lmsg.routing_id = 0; + _u.lmsg.content = reinterpret_cast ( + alloc_->allocate (size_ + sizeof (content_t))); + + if (!_u.lmsg.content) { + errno = ENOMEM; + return -1; + } - _u.lmsg.content->data = _u.lmsg.content + 1; - _u.lmsg.content->size = size_; - _u.lmsg.content->ffn = (zmq_free_fn *) alloc_->deallocate_msg; - _u.lmsg.content->hint = alloc_; - new (&_u.lmsg.content->refcnt) zmq::atomic_counter_t (); + _u.lmsg.content->data = _u.lmsg.content + 1; + _u.lmsg.content->size = size_; + _u.lmsg.content->ffn = (zmq_free_fn *) alloc_->deallocate_msg; + _u.lmsg.content->hint = alloc_; + new (&_u.lmsg.content->refcnt) zmq::atomic_counter_t (); + } return 0; }