From b307c9765c6fa8d88d8c2fcf9de44eec5a7765d6 Mon Sep 17 00:00:00 2001 From: "yinzefeng.yzf" Date: Mon, 2 Feb 2026 22:25:18 +0800 Subject: [PATCH 01/11] add buffer pool & spsc_queue --- src/ailego/buffer/buffer_manager.cc | 1 + src/include/zvec/ailego/buffer/buffer_pool.h | 311 +++++++++++++++++++ 2 files changed, 312 insertions(+) create mode 100644 src/include/zvec/ailego/buffer/buffer_pool.h diff --git a/src/ailego/buffer/buffer_manager.cc b/src/ailego/buffer/buffer_manager.cc index ac2945b0..307e80ce 100644 --- a/src/ailego/buffer/buffer_manager.cc +++ b/src/ailego/buffer/buffer_manager.cc @@ -20,6 +20,7 @@ #include #include #include +#include #ifdef __clang__ #pragma clang diagnostic push diff --git a/src/include/zvec/ailego/buffer/buffer_pool.h b/src/include/zvec/ailego/buffer/buffer_pool.h new file mode 100644 index 00000000..5a09abfa --- /dev/null +++ b/src/include/zvec/ailego/buffer/buffer_pool.h @@ -0,0 +1,311 @@ +#pragma once + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +using block_id_t = int; + +#define BLOCK_SIZE (4 * 1024 * 1024) // 2 MB +#define BLOCK_MASK (BLOCK_SIZE - 1) +#define BLOCK_ID(offset) (offset >> 22) +#define BLOCK_OFFSET(offset) (offset & BLOCK_MASK) + +class LRUCache { + boost::lockfree::spsc_queue> q; +}; + +class LPMap { + struct Entry { + std::atomic ref_count; + char* buffer; + }; + + public: + LPMap() : entry_num_(0), entries_(nullptr) {} + ~LPMap() { + delete[] entries_; + } + + void init(size_t entry_num) { + if (entries_) { + delete[] entries_; + } + entry_num_ = entry_num; + entries_ = new Entry[entry_num_]; + for (size_t i = 0; i < entry_num_; i++) { + // entries_[i].ref_count.store(0); + entries_[i].ref_count.store(std::numeric_limits::min()); + entries_[i].buffer = nullptr; + } + } + + char* acquire_block(block_id_t block_id) { + assert(block_id < entry_num_); + Entry& entry = entries_[block_id]; + int rc = entry.ref_count.fetch_add(1); + if (rc < 0) { + return nullptr; + } + return entry.buffer; + } + + void release_block(block_id_t block_id) { + assert(block_id < entry_num_); + Entry& entry = entries_[block_id]; + int rc = entry.ref_count.fetch_sub(1); + assert(rc > 0); + } + + // need be called under lock + char* evict_block(block_id_t block_id) { + assert(block_id < entry_num_); + Entry& entry = entries_[block_id]; + int expected = 0; + if (entry.ref_count.compare_exchange_strong(expected, std::numeric_limits::min())) { + char* buffer = entry.buffer; + entry.buffer = nullptr; + return buffer; + } else { + return nullptr; + } + } + + // need be called under lock + char* set_block_acquired(block_id_t block_id, char* buffer) { + // std::cout << "Set block " << block_id << std::endl; + assert(block_id < entry_num_); + Entry& entry = entries_[block_id]; + if (entry.ref_count.load() >= 0) { + entry.ref_count.fetch_add(1); + return entry.buffer; + } + entry.buffer = buffer; + entry.ref_count.store(1); + return buffer; + } + + // need be called under lock + void recycle(std::queue& free_buffers) { + for (size_t i = 0; i < entry_num_; i++) { + Entry& entry = entries_[i]; + if (entry.ref_count.load() == 0) { + char* buffer = evict_block(i); + if (buffer) { + free_buffers.push(buffer); + } + } + } + } + + size_t entry_num() const { + return entry_num_; + } + + private: + Entry* entries_; + size_t entry_num_; +}; + +class BufferPool; + +struct BufferPoolHandle { + BufferPoolHandle(BufferPool& pool); + BufferPoolHandle(BufferPoolHandle&& other) : pool(other.pool), local_cache(std::move(other.local_cache)), hit_num_(other.hit_num_) { + other.local_cache.clear(); + other.hit_num_ = 0; + } + ~BufferPoolHandle(); + + char* get_block(size_t offset, size_t size); + + void release_all(); + + BufferPool& pool; +#ifdef USE_LOCAL_CACHE + // std::unordered_map local_cache; + phmap::flat_hash_map local_cache; +#else + std::vector local_cache; +#endif + int hit_num_; +}; + +class BufferPool { + public: + BufferPool(const std::string& filename, size_t pool_capacity) : pool_capacity_(pool_capacity){ + fd_ = open(filename.c_str(), O_RDONLY); + if (fd_ < 0) { + throw std::runtime_error("Failed to open file: " + filename); + } + struct stat st; + if (fstat(fd_, &st) < 0) { + throw std::runtime_error("Failed to stat file: " + filename); + } + file_size_ = st.st_size; + lp_map_.init((file_size_ + BLOCK_SIZE - 1) / BLOCK_SIZE); + + size_t buffer_num = pool_capacity_ / BLOCK_SIZE; + for (size_t i = 0; i < buffer_num; i++) { + char* buffer = (char*)aligned_alloc(64, BLOCK_SIZE); + free_buffers_.push(buffer); + } + std::cout << "buffer_num: " << buffer_num << std::endl; + std::cout << "entry_num: " << lp_map_.entry_num() << std::endl; + } + ~BufferPool() { + close(fd_); + } + + BufferPoolHandle get_handle() { + return BufferPoolHandle(*this); + } + + char* acquire_buffer(block_id_t block_id, int retry = 0) { + char* buffer = lp_map_.acquire_block(block_id); + if (buffer) { + return buffer; + } + { + std::lock_guard lock(mutex_); + if (free_buffers_.empty()) { + for (int i = 0; i < retry; i++) { + lp_map_.recycle(free_buffers_); + if (!free_buffers_.empty()) { + break; + } + } + } + if (free_buffers_.empty()) { + return nullptr; + } + buffer = free_buffers_.front(); + free_buffers_.pop(); + } + size_t read_offset = static_cast(block_id) * BLOCK_SIZE; + size_t to_read = std::min(BLOCK_SIZE, file_size_ - read_offset); + + ssize_t read_bytes = pread(fd_, buffer, to_read, read_offset); + if (read_bytes != static_cast(to_read)) { + std::cerr << "Failed to read file at offset " << read_offset << std::endl; + exit(-1); + } + + { + std::lock_guard lock(mutex_); + char* placed_buffer = lp_map_.set_block_acquired(block_id, buffer); + if (placed_buffer != buffer) { + // another thread has set the block + free_buffers_.push(buffer); + } + return placed_buffer; + } + } + + size_t file_size() const { + return file_size_; + } + + private: + int fd_; + size_t file_size_; + size_t pool_capacity_; + + public: + LPMap lp_map_; + + private: + std::mutex mutex_; + std::queue free_buffers_; +}; + + +struct Counter { + ~Counter() = default; + + static Counter& get_instance() { + static Counter instance; + return instance; + } + + void record(const std::string& name, int64_t value) { + auto it = static_counters.find(name); + if (it == static_counters.end()) { + auto counter = std::make_unique>(0); + it = static_counters.emplace(name, std::move(counter)).first; + } + it->second->fetch_add(value); + } + + void display() { + for (const auto& pair : static_counters) { + std::cout << pair.first << ": " << pair.second->load() << std::endl; + } + } + + void clear() { + static_counters.clear(); + } + + private: + Counter() {} + std::map>> static_counters; +}; + +BufferPoolHandle::BufferPoolHandle(BufferPool& pool) : pool(pool), hit_num_(0) {} +BufferPoolHandle::~BufferPoolHandle() { + Counter::get_instance().record("buffer_pool_handle_hit_num", hit_num_); + release_all(); +} + +char* BufferPoolHandle::get_block(size_t offset, size_t size) { + block_id_t block_id = BLOCK_ID(offset); + assert(block_id == BLOCK_ID(offset + size - 1)); +#ifdef USE_LOCAL_CACHE + auto it = local_cache.find(block_id); + if (it != local_cache.end()) { + hit_num_++; + return it->second + BLOCK_OFFSET(offset); + } +#endif + + char* buffer = pool.acquire_buffer(block_id, 3); + if (buffer) { +#ifdef USE_LOCAL_CACHE + local_cache[block_id] = buffer; +#else + local_cache.push_back(block_id); +#endif + return buffer + BLOCK_OFFSET(offset); + } + + return nullptr; +} + +void BufferPoolHandle::release_all() { +#ifdef USE_LOCAL_CACHE + Counter::get_instance().record("buffer_pool_handle_release_call", local_cache.size()); + for (const auto& pair : local_cache) { + pool.lp_map_.release_block(pair.first); + } +#else + for (block_id_t block_id : local_cache) { + pool.lp_map_.release_block(block_id); + } +#endif + local_cache.clear(); +} \ No newline at end of file From a96e684d276767ff4fc0c6019cf923d9f2707080 Mon Sep 17 00:00:00 2001 From: "yinzefeng.yzf" Date: Wed, 4 Feb 2026 16:09:11 +0800 Subject: [PATCH 02/11] add buffer pool & open buffer storage ut --- src/ailego/buffer/buffer_manager.cc | 1 - src/core/utility/buffer1_storage.cc | 438 ++ src/core/utility/buffer_storage.cc | 2 +- src/include/zvec/ailego/buffer/buffer_pool.h | 520 ++- .../zvec/ailego/buffer/concurrentqueue.h | 3747 +++++++++++++++++ ..._test.cpp => flat_streamer_buffer_test.cc} | 0 ....cpp => flat_streamer_buffer_time_test.cc} | 0 7 files changed, 4474 insertions(+), 234 deletions(-) create mode 100644 src/core/utility/buffer1_storage.cc create mode 100644 src/include/zvec/ailego/buffer/concurrentqueue.h rename tests/core/algorithm/flat/{flat_streamer_buffer_test.cpp => flat_streamer_buffer_test.cc} (100%) rename tests/core/algorithm/flat/{flat_streamer_buffer_time_test.cpp => flat_streamer_buffer_time_test.cc} (100%) diff --git a/src/ailego/buffer/buffer_manager.cc b/src/ailego/buffer/buffer_manager.cc index 307e80ce..ac2945b0 100644 --- a/src/ailego/buffer/buffer_manager.cc +++ b/src/ailego/buffer/buffer_manager.cc @@ -20,7 +20,6 @@ #include #include #include -#include #ifdef __clang__ #pragma clang diagnostic push diff --git a/src/core/utility/buffer1_storage.cc b/src/core/utility/buffer1_storage.cc new file mode 100644 index 00000000..0ea591d9 --- /dev/null +++ b/src/core/utility/buffer1_storage.cc @@ -0,0 +1,438 @@ +// Copyright 2025-present the zvec project +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include +// #include +#include +#include +#include +#include +#include +#include "utility_params.h" + +#include + +namespace zvec { +namespace core { + +/*! MMap File Storage + */ +class Buffer1Storage : public IndexStorage { + public: + /*! Index Storage Segment + */ + class Segment : public IndexStorage::Segment, + public std::enable_shared_from_this { + public: + //! Index Storage Pointer + typedef std::shared_ptr Pointer; + + //! Constructor + Segment(Buffer1Storage *owner, IndexMapping::Segment *segment, size_t segment_id) + : segment_(segment), + owner_(owner), + segment_id_(segment_id), + capacity_(static_cast(segment->meta()->data_size + + segment->meta()->padding_size)) {} + + //! Destructor + virtual ~Segment(void) {} + + //! Retrieve size of data + size_t data_size(void) const override { + return static_cast(segment_->meta()->data_size); + } + + //! Retrieve crc of data + uint32_t data_crc(void) const override { + return segment_->meta()->data_crc; + } + + //! Retrieve size of padding + size_t padding_size(void) const override { + return static_cast(segment_->meta()->padding_size); + } + + //! Retrieve capacity of segment + size_t capacity(void) const override { + return capacity_; + } + + //! Fetch data from segment (with own buffer) + size_t fetch(size_t offset, void *buf, size_t len) const override { + if (ailego_unlikely(offset + len > segment_->meta()->data_size)) { + auto meta = segment_->meta(); + if (offset > meta->data_size) { + offset = meta->data_size; + } + len = meta->data_size - offset; + } + memmove(buf, (const uint8_t *)(owner_->get_buffer(offset, len, segment_id_)) + offset, + len); + return len; + } + + //! Read data from segment + size_t read(size_t offset, const void **data, size_t len) override { + + if (ailego_unlikely(offset + len > segment_->meta()->data_size)) { + auto meta = segment_->meta(); + if (offset > meta->data_size) { + offset = meta->data_size; + } + len = meta->data_size - offset; + } + size_t segment_offset = segment_->meta()->data_index + owner_->get_context_offset(); + *data = owner_->get_buffer(segment_offset, capacity_, segment_id_) + offset; + return len; + } + + size_t read(size_t offset, MemoryBlock &data, size_t len) override { + if (ailego_unlikely(offset + len > segment_->meta()->data_size)) { + auto meta = segment_->meta(); + if (offset > meta->data_size) { + offset = meta->data_size; + } + len = meta->data_size - offset; + } + size_t segment_offset = segment_->meta()->data_index + owner_->get_context_offset(); + data.reset(owner_->get_buffer(segment_offset, capacity_, segment_id_) + offset); + if (data.data()) { + return len; + } else { + LOG_ERROR("read error."); + return -1; + } + } + + //! Write data into the storage with offset + size_t write(size_t /*offset*/, const void * /*data*/, + size_t len) override { + return len; + } + + //! Resize size of data + size_t resize(size_t /*size*/) override { + return 0; + } + + //! Update crc of data + void update_data_crc(uint32_t /*crc*/) override {} + + //! Clone the segment + IndexStorage::Segment::Pointer clone(void) override { + return shared_from_this(); + } + + private: + IndexMapping::Segment *segment_{}; + Buffer1Storage *owner_{nullptr}; + size_t capacity_{}; + size_t segment_id_{}; + }; + + //! Destructor + virtual ~Buffer1Storage(void) { + this->cleanup(); + } + + //! Initialize storage + int init(const ailego::Params & /*params*/) override { + return 0; + } + + //! Cleanup storage + int cleanup(void) override { + this->close_index(); + return 0; + } + + //! Open storage + int open(const std::string &path, bool /*create*/) override { + LOG_INFO("open buffer storage 1"); + file_name_ = path; + buffer_pool_ = std::make_unique(path, 10u * 1024 * 1024 * 1024, 2490368 * 2); + buffer_pool_handle_ = + std::make_unique(buffer_pool_->get_handle()); + int ret = ParseToMapping(); + LOG_ERROR("segment count: %lu, max_segment_size: %lu", segments_.size(), max_segment_size_); + if(ret != 0) { + return ret; + } + return 0; + } + + char *get_buffer(size_t offset, size_t length, size_t block_id) { + return buffer_pool_handle_->get_block(offset, length, block_id); + } + + int get_meta(size_t offset, size_t length, char *out) { + return buffer_pool_handle_->get_meta(offset, length, out); + } + + int ParseHeader(size_t offset) { + char *buffer = new char[sizeof(header_)]; + get_meta(offset, sizeof(header_), buffer); + uint8_t *header_ptr = reinterpret_cast(buffer); + memcpy(&header_, header_ptr, sizeof(header_)); + delete[] buffer; + if (header_.meta_header_size != sizeof(IndexFormat::MetaHeader)) { + LOG_ERROR("Header meta size is invalid."); + return IndexError_InvalidLength; + } + if (ailego::Crc32c::Hash(&header_, sizeof(header_), header_.header_crc) != + header_.header_crc) { + LOG_ERROR("Header meta checksum is invalid."); + return IndexError_InvalidChecksum; + } + return 0; + } + + int ParseFooter(size_t offset) { + char *buffer = new char[sizeof(footer_)]; + get_meta(offset, sizeof(footer_), buffer); + uint8_t *footer_ptr = reinterpret_cast(buffer); + memcpy(&footer_, footer_ptr, sizeof(footer_)); + delete[] buffer; + if (offset < (size_t)footer_.segments_meta_size) { + LOG_ERROR("Footer meta size is invalid."); + return IndexError_InvalidLength; + } + if (ailego::Crc32c::Hash(&footer_, sizeof(footer_), footer_.footer_crc) != + footer_.footer_crc) { + LOG_ERROR("Footer meta checksum is invalid."); + return IndexError_InvalidChecksum; + } + return 0; + } + + int ParseSegment(size_t offset) { + segment_buffer_ = std::make_unique(footer_.segments_meta_size); + get_meta(offset, footer_.segments_meta_size, segment_buffer_.get()); + if (ailego::Crc32c::Hash(segment_buffer_.get(), footer_.segments_meta_size, 0u) != + footer_.segments_meta_crc) { + LOG_ERROR("Index segments meta checksum is invalid."); + return IndexError_InvalidChecksum; + } + IndexFormat::SegmentMeta *segment_start = + reinterpret_cast(segment_buffer_.get()); + uint32_t segment_ids_offset = footer_.segments_meta_size; + for (IndexFormat::SegmentMeta *iter = segment_start, + *end = segment_start + footer_.segment_count; + iter != end; ++iter) { + if (iter->segment_id_offset > footer_.segments_meta_size) { + return IndexError_InvalidValue; + } + if (iter->data_index > footer_.content_size) { + return IndexError_InvalidValue; + } + if (iter->data_index + iter->data_size > footer_.content_size) { + return IndexError_InvalidLength; + } + + if (iter->segment_id_offset < segment_ids_offset) { + segment_ids_offset = iter->segment_id_offset; + } + id_hash_.emplace( + std::string(reinterpret_cast(segment_start) + + iter->segment_id_offset), + segments_.size()); + segments_.emplace( + std::string(reinterpret_cast(segment_start) + + iter->segment_id_offset), + iter); + max_segment_size_ = std::max(max_segment_size_, iter->data_size + iter->padding_size); + if (sizeof(IndexFormat::SegmentMeta) * footer_.segment_count > + footer_.segments_meta_size) { + return IndexError_InvalidLength; + } + } + return 0; + } + + int ParseToMapping() { + ParseHeader(0); + // Unpack footer + if (header_.meta_footer_size != sizeof(IndexFormat::MetaFooter)) { + return IndexError_InvalidLength; + } + if ((int32_t)header_.meta_footer_offset < 0) { + return IndexError_Unsupported; + } + size_t footer_offset = header_.meta_footer_offset; + ParseFooter(footer_offset); + + // Unpack segment table + if (sizeof(IndexFormat::SegmentMeta) * footer_.segment_count > + footer_.segments_meta_size) { + return IndexError_InvalidLength; + } + const size_t segment_start_offset = footer_offset - footer_.segments_meta_size; + ParseSegment(segment_start_offset); + return 0; + } + + //! Flush storage + int flush(void) override { + return this->flush_index(); + } + + //! Close storage + int close(void) override { + this->close_index(); + return 0; + } + + //! Append a segment into storage + int append(const std::string &id, size_t size) override { + return this->append_segment(id, size); + } + + //! Refresh meta information (checksum, update time, etc.) + void refresh(uint64_t chkp) override { + this->refresh_index(chkp); + } + + //! Retrieve check point of storage + uint64_t check_point(void) const override { + return footer_.check_point; + } + + //! Retrieve a segment by id + IndexStorage::Segment::Pointer get(const std::string &id, int) override { + IndexMapping::Segment *segment = this->get_segment(id); + if (!segment) { + return Buffer1Storage::Segment::Pointer(); + } + return std::make_shared(this, segment, + id_hash_[id]); + } + + //! Test if it a segment exists + bool has(const std::string &id) const override { + return this->has_segment(id); + } + + //! Retrieve magic number of index + uint32_t magic(void) const override { + return header_.magic; + } + + uint32_t get_context_offset() { + return header_.content_offset; + } + + protected: + //! Initialize index version segment + int init_version_segment(void) { + size_t data_size = std::strlen(IndexVersion::Details()); + int error_code = + this->append_segment(INDEX_VERSION_SEGMENT_NAME, data_size); + if (error_code != 0) { + return error_code; + } + + IndexMapping::Segment *segment = get_segment(INDEX_VERSION_SEGMENT_NAME); + if (!segment) { + return IndexError_MMapFile; + } + auto meta = segment->meta(); + size_t capacity = static_cast(meta->padding_size + meta->data_size); + memcpy(segment->data(), IndexVersion::Details(), data_size); + segment->set_dirty(); + meta->data_crc = ailego::Crc32c::Hash(segment->data(), data_size, 0); + meta->data_size = data_size; + meta->padding_size = capacity - data_size; + return 0; + } + + //! Initialize index file + int init_index(const std::string &path) { + // Add index version + int error_code = this->init_version_segment(); + if (error_code != 0) { + return error_code; + } + + // Refresh mapping + this->refresh_index(0); + return 0; + } + + //! Set the index file as dirty + void set_as_dirty(void) { + index_dirty_ = true; + } + + //! Refresh meta information (checksum, update time, etc.) + void refresh_index(uint64_t /*chkp*/) {} + + //! Flush index storage + int flush_index(void) { + return 0; + } + + //! Close index storage + void close_index(void) { + std::lock_guard latch(mapping_mutex_); + file_name_.clear(); + segments_.clear(); + memset(&header_, 0, sizeof(header_)); + memset(&footer_, 0, sizeof(footer_)); + segment_buffer_.release(); + } + + //! Append a segment into storage + int append_segment(const std::string & /*id*/, size_t /*size*/) { + return 0; + } + + //! Test if a segment exists + bool has_segment(const std::string &id) const { + std::lock_guard latch(mapping_mutex_); + return (segments_.find(id) != segments_.end()); + } + + //! Get a segment from storage + IndexMapping::Segment *get_segment(const std::string &id) { + std::lock_guard latch(mapping_mutex_); + auto iter = segments_.find(id); + if (iter == segments_.end()) { + return nullptr; + } + IndexMapping::Segment *item = &iter->second; + return item; + } + + private: + bool index_dirty_{false}; + mutable std::mutex mapping_mutex_{}; + + // buffer manager + std::string file_name_; + IndexFormat::MetaHeader header_; + IndexFormat::MetaFooter footer_; + std::map segments_{}; + std::map id_hash_{}; + size_t max_segment_size_{0}; + std::unique_ptr segment_buffer_{nullptr}; + + std::unique_ptr buffer_pool_{nullptr}; + std::unique_ptr buffer_pool_handle_{nullptr}; +}; + +INDEX_FACTORY_REGISTER_STORAGE_ALIAS(BufferStorage, Buffer1Storage); + +} // namespace core +} // namespace zvec \ No newline at end of file diff --git a/src/core/utility/buffer_storage.cc b/src/core/utility/buffer_storage.cc index 4ac3c6b3..d4b23c87 100644 --- a/src/core/utility/buffer_storage.cc +++ b/src/core/utility/buffer_storage.cc @@ -436,7 +436,7 @@ class BufferStorage : public IndexStorage { std::map segments_{}; }; -INDEX_FACTORY_REGISTER_STORAGE(BufferStorage); +// INDEX_FACTORY_REGISTER_STORAGE(BufferStorage); } // namespace core } // namespace zvec diff --git a/src/include/zvec/ailego/buffer/buffer_pool.h b/src/include/zvec/ailego/buffer/buffer_pool.h index 5a09abfa..d86cffec 100644 --- a/src/include/zvec/ailego/buffer/buffer_pool.h +++ b/src/include/zvec/ailego/buffer/buffer_pool.h @@ -1,311 +1,367 @@ #pragma once +#include +#include +#include #include #include #include #include #include +#include +#include +#include #include #include +#include #include #include -#include -#include -#include -#include -#include -#include -#include -#include +#include "concurrentqueue.h" using block_id_t = int; - -#define BLOCK_SIZE (4 * 1024 * 1024) // 2 MB -#define BLOCK_MASK (BLOCK_SIZE - 1) -#define BLOCK_ID(offset) (offset >> 22) -#define BLOCK_OFFSET(offset) (offset & BLOCK_MASK) +using version_t = int; class LRUCache { - boost::lockfree::spsc_queue> q; -}; - -class LPMap { - struct Entry { - std::atomic ref_count; - char* buffer; - }; - public: - LPMap() : entry_num_(0), entries_(nullptr) {} - ~LPMap() { - delete[] entries_; + typedef std::pair BlockType; + typedef moodycamel::ConcurrentQueue ConcurrentQueue; + + int init(size_t block_size) { + for(int i = 0; i < CATCH_QUEUE_NUM; i++) { + queues_.push_back(ConcurrentQueue(block_size)); + } + return 0; } - void init(size_t entry_num) { - if (entries_) { - delete[] entries_; - } - entry_num_ = entry_num; - entries_ = new Entry[entry_num_]; - for (size_t i = 0; i < entry_num_; i++) { - // entries_[i].ref_count.store(0); - entries_[i].ref_count.store(std::numeric_limits::min()); - entries_[i].buffer = nullptr; + BlockType evict_single_block() { + BlockType item; + for(int i = 0; i < CATCH_QUEUE_NUM; i++) { + bool found = queues_[i].try_dequeue(item); + if(found) { + break; } + } + return item; } - char* acquire_block(block_id_t block_id) { - assert(block_id < entry_num_); - Entry& entry = entries_[block_id]; - int rc = entry.ref_count.fetch_add(1); - if (rc < 0) { - return nullptr; - } - return entry.buffer; + bool add_single_block(const BlockType &block, int block_type) { + std::cout << "in LRU: " << block.first << ", " << block.second << std::endl; + return queues_[block_type].try_enqueue(block); } - void release_block(block_id_t block_id) { - assert(block_id < entry_num_); - Entry& entry = entries_[block_id]; - int rc = entry.ref_count.fetch_sub(1); - assert(rc > 0); - } + private: + constexpr static size_t CATCH_QUEUE_NUM = 3; + std::vector queues_; +}; - // need be called under lock - char* evict_block(block_id_t block_id) { - assert(block_id < entry_num_); - Entry& entry = entries_[block_id]; - int expected = 0; - if (entry.ref_count.compare_exchange_strong(expected, std::numeric_limits::min())) { - char* buffer = entry.buffer; - entry.buffer = nullptr; - return buffer; - } else { - return nullptr; - } +class LPMap { + struct Entry { + std::atomic ref_count; + std::atomic load_count; + char *buffer; + }; + + public: + LPMap() : entry_num_(0), entries_(nullptr) {} + ~LPMap() { + delete[] entries_; + } + + void init(size_t entry_num) { + if (entries_) { + delete[] entries_; } - - // need be called under lock - char* set_block_acquired(block_id_t block_id, char* buffer) { - // std::cout << "Set block " << block_id << std::endl; - assert(block_id < entry_num_); - Entry& entry = entries_[block_id]; - if (entry.ref_count.load() >= 0) { - entry.ref_count.fetch_add(1); - return entry.buffer; - } - entry.buffer = buffer; - entry.ref_count.store(1); - return buffer; + entry_num_ = entry_num; + entries_ = new Entry[entry_num_]; + for (size_t i = 0; i < entry_num_; i++) { + entries_[i].ref_count.store(std::numeric_limits::min()); + entries_[i].load_count.store(0); + entries_[i].buffer = nullptr; } - - // need be called under lock - void recycle(std::queue& free_buffers) { - for (size_t i = 0; i < entry_num_; i++) { - Entry& entry = entries_[i]; - if (entry.ref_count.load() == 0) { - char* buffer = evict_block(i); - if (buffer) { - free_buffers.push(buffer); - } - } - } + } + + char *acquire_block(block_id_t block_id) { + assert(block_id < entry_num_); + Entry &entry = entries_[block_id]; + int rc = entry.ref_count.fetch_add(1); + if (rc < 0) { + return nullptr; } - - size_t entry_num() const { - return entry_num_; + return entry.buffer; + } + + void release_block(block_id_t block_id) { + assert(block_id < entry_num_); + Entry &entry = entries_[block_id]; + int rc = entry.ref_count.fetch_sub(1); + assert(rc >= 0); + if(rc == 0) { + LRUCache::BlockType block; + block.first = block_id; + block.second = entry.load_count.load(); + cache_.add_single_block(block, 0); + } + } + + // need be called under lock + char *evict_block(block_id_t block_id) { + assert(block_id < entry_num_); + Entry &entry = entries_[block_id]; + int expected = 0; + if (entry.ref_count.compare_exchange_strong( + expected, std::numeric_limits::min())) { + char *buffer = entry.buffer; + entry.buffer = nullptr; + return buffer; + } else { + return nullptr; } + } + + // need be called under lock + char *set_block_acquired(block_id_t block_id, char *buffer) { + // std::cout << "Set block " << block_id << std::endl; + assert(block_id < entry_num_); + Entry &entry = entries_[block_id]; + if (entry.ref_count.load() >= 0) { + entry.ref_count.fetch_add(1); + return entry.buffer; + } + entry.buffer = buffer; + entry.ref_count.store(1); + entry.load_count.fetch_add(1); + return buffer; + } + + // need be called under lock + void recycle(std::queue &free_buffers) { + LRUCache::BlockType block; + do { + block = cache_.evict_single_block(); + } while(isDeadBlock(block)); + char *buffer = evict_block(block.first); + if (buffer) { + free_buffers.push(buffer); + } + } - private: - Entry* entries_; - size_t entry_num_; + size_t entry_num() const { + return entry_num_; + } + + private: + Entry *entries_; + size_t entry_num_; + LRUCache cache_; + + bool isDeadBlock(LRUCache::BlockType block) { + Entry &entry = entries_[block.first]; + return block.second == entry.load_count.load(); + } }; -class BufferPool; +class VecBufferPool; -struct BufferPoolHandle { - BufferPoolHandle(BufferPool& pool); - BufferPoolHandle(BufferPoolHandle&& other) : pool(other.pool), local_cache(std::move(other.local_cache)), hit_num_(other.hit_num_) { - other.local_cache.clear(); - other.hit_num_ = 0; - } - ~BufferPoolHandle(); +struct VecBufferPoolHandle { + VecBufferPoolHandle(VecBufferPool &pool); + VecBufferPoolHandle(VecBufferPoolHandle &&other) + : pool(other.pool), + local_cache(std::move(other.local_cache)), + hit_num_(other.hit_num_) { + other.local_cache.clear(); + other.hit_num_ = 0; + } + ~VecBufferPoolHandle(); - char* get_block(size_t offset, size_t size); + char *get_block(size_t offset, size_t size, size_t block_id); - void release_all(); + int get_meta(size_t offset, size_t length, char *buffer); - BufferPool& pool; + void release_all(); + + VecBufferPool &pool; #ifdef USE_LOCAL_CACHE - // std::unordered_map local_cache; - phmap::flat_hash_map local_cache; + // std::unordered_map local_cache; + phmap::flat_hash_map local_cache; #else - std::vector local_cache; + std::vector local_cache; #endif - int hit_num_; + int hit_num_; }; -class BufferPool { - public: - BufferPool(const std::string& filename, size_t pool_capacity) : pool_capacity_(pool_capacity){ - fd_ = open(filename.c_str(), O_RDONLY); - if (fd_ < 0) { - throw std::runtime_error("Failed to open file: " + filename); - } - struct stat st; - if (fstat(fd_, &st) < 0) { - throw std::runtime_error("Failed to stat file: " + filename); - } - file_size_ = st.st_size; - lp_map_.init((file_size_ + BLOCK_SIZE - 1) / BLOCK_SIZE); - - size_t buffer_num = pool_capacity_ / BLOCK_SIZE; - for (size_t i = 0; i < buffer_num; i++) { - char* buffer = (char*)aligned_alloc(64, BLOCK_SIZE); - free_buffers_.push(buffer); - } - std::cout << "buffer_num: " << buffer_num << std::endl; - std::cout << "entry_num: " << lp_map_.entry_num() << std::endl; +class VecBufferPool { + public: + VecBufferPool(const std::string &filename, size_t pool_capacity, size_t block_size) + : pool_capacity_(pool_capacity) { + fd_ = open(filename.c_str(), O_RDONLY); + if (fd_ < 0) { + throw std::runtime_error("Failed to open file: " + filename); } - ~BufferPool() { - close(fd_); + struct stat st; + if (fstat(fd_, &st) < 0) { + throw std::runtime_error("Failed to stat file: " + filename); } + file_size_ = st.st_size; - BufferPoolHandle get_handle() { - return BufferPoolHandle(*this); + size_t buffer_num = pool_capacity_ / block_size; + lp_map_.init(buffer_num); + for (size_t i = 0; i < buffer_num; i++) { + char *buffer = (char *)aligned_alloc(64, block_size); + free_buffers_.push(buffer); } - - char* acquire_buffer(block_id_t block_id, int retry = 0) { - char* buffer = lp_map_.acquire_block(block_id); - if (buffer) { - return buffer; - } - { - std::lock_guard lock(mutex_); - if (free_buffers_.empty()) { - for (int i = 0; i < retry; i++) { - lp_map_.recycle(free_buffers_); - if (!free_buffers_.empty()) { - break; - } - } - } - if (free_buffers_.empty()) { - return nullptr; - } - buffer = free_buffers_.front(); - free_buffers_.pop(); + std::cout << "buffer_num: " << buffer_num << std::endl; + std::cout << "entry_num: " << lp_map_.entry_num() << std::endl; + } + ~VecBufferPool() { + close(fd_); + } + + VecBufferPoolHandle get_handle() { + return VecBufferPoolHandle(*this); + } + + char *acquire_buffer(block_id_t block_id, size_t offset, size_t size, int retry = 0) { + char *buffer = lp_map_.acquire_block(block_id); + if (buffer) { + return buffer; + } + { + std::lock_guard lock(mutex_); + if (free_buffers_.empty()) { + for (int i = 0; i < retry; i++) { + lp_map_.recycle(free_buffers_); + if (!free_buffers_.empty()) { + break; + } } - size_t read_offset = static_cast(block_id) * BLOCK_SIZE; - size_t to_read = std::min(BLOCK_SIZE, file_size_ - read_offset); + } + if (free_buffers_.empty()) { + return nullptr; + } + buffer = free_buffers_.front(); + free_buffers_.pop(); + } - ssize_t read_bytes = pread(fd_, buffer, to_read, read_offset); - if (read_bytes != static_cast(to_read)) { - std::cerr << "Failed to read file at offset " << read_offset << std::endl; - exit(-1); - } + ssize_t read_bytes = pread(fd_, buffer, size, offset); + if (read_bytes != static_cast(size)) { + std::cerr << "Failed to read file at offset " << offset << std::endl; + exit(-1); + } - { - std::lock_guard lock(mutex_); - char* placed_buffer = lp_map_.set_block_acquired(block_id, buffer); - if (placed_buffer != buffer) { - // another thread has set the block - free_buffers_.push(buffer); - } - return placed_buffer; - } + { + std::lock_guard lock(mutex_); + char *placed_buffer = lp_map_.set_block_acquired(block_id, buffer); + if (placed_buffer != buffer) { + // another thread has set the block + free_buffers_.push(buffer); + } + return placed_buffer; } + } - size_t file_size() const { - return file_size_; + int get_meta(size_t offset, size_t length, char *buffer) { + ssize_t read_bytes = pread(fd_, buffer, length, offset); + if (read_bytes != static_cast(length)) { + std::cerr << "Failed to read file at offset " << offset << std::endl; + exit(-1); } + return 0; + } - private: - int fd_; - size_t file_size_; - size_t pool_capacity_; + size_t file_size() const { + return file_size_; + } - public: - LPMap lp_map_; + private: + int fd_; + size_t file_size_; + size_t pool_capacity_; - private: - std::mutex mutex_; - std::queue free_buffers_; + public: + LPMap lp_map_; + + private: + std::mutex mutex_; + std::queue free_buffers_; }; struct Counter { - ~Counter() = default; - - static Counter& get_instance() { - static Counter instance; - return instance; + ~Counter() = default; + + static Counter &get_instance() { + static Counter instance; + return instance; + } + + void record(const std::string &name, int64_t value) { + auto it = static_counters.find(name); + if (it == static_counters.end()) { + auto counter = std::make_unique>(0); + it = static_counters.emplace(name, std::move(counter)).first; } + it->second->fetch_add(value); + } - void record(const std::string& name, int64_t value) { - auto it = static_counters.find(name); - if (it == static_counters.end()) { - auto counter = std::make_unique>(0); - it = static_counters.emplace(name, std::move(counter)).first; - } - it->second->fetch_add(value); + void display() { + for (const auto &pair : static_counters) { + std::cout << pair.first << ": " << pair.second->load() << std::endl; } + } - void display() { - for (const auto& pair : static_counters) { - std::cout << pair.first << ": " << pair.second->load() << std::endl; - } - } - - void clear() { - static_counters.clear(); - } + void clear() { + static_counters.clear(); + } - private: - Counter() {} - std::map>> static_counters; + private: + Counter() {} + std::map>> static_counters; }; -BufferPoolHandle::BufferPoolHandle(BufferPool& pool) : pool(pool), hit_num_(0) {} -BufferPoolHandle::~BufferPoolHandle() { - Counter::get_instance().record("buffer_pool_handle_hit_num", hit_num_); - release_all(); +VecBufferPoolHandle::VecBufferPoolHandle(VecBufferPool &pool) + : pool(pool), hit_num_(0) {} +VecBufferPoolHandle::~VecBufferPoolHandle() { + Counter::get_instance().record("buffer_pool_handle_hit_num", hit_num_); + release_all(); } -char* BufferPoolHandle::get_block(size_t offset, size_t size) { - block_id_t block_id = BLOCK_ID(offset); - assert(block_id == BLOCK_ID(offset + size - 1)); +char *VecBufferPoolHandle::get_block(size_t offset, size_t size, size_t block_id) { #ifdef USE_LOCAL_CACHE - auto it = local_cache.find(block_id); - if (it != local_cache.end()) { - hit_num_++; - return it->second + BLOCK_OFFSET(offset); - } + auto it = local_cache.find(block_id); + if (it != local_cache.end()) { + hit_num_++; + return it->second; + } #endif - char* buffer = pool.acquire_buffer(block_id, 3); - if (buffer) { + char *buffer = pool.acquire_buffer(block_id, offset, size, 3); + if (buffer) { #ifdef USE_LOCAL_CACHE - local_cache[block_id] = buffer; + local_cache[block_id] = buffer; #else - local_cache.push_back(block_id); + local_cache.push_back(block_id); #endif - return buffer + BLOCK_OFFSET(offset); - } + return buffer; + } - return nullptr; + return nullptr; } -void BufferPoolHandle::release_all() { +int VecBufferPoolHandle::get_meta(size_t offset, size_t length, char *out) { + return pool.get_meta(offset, length, out); +} + +void VecBufferPoolHandle::release_all() { #ifdef USE_LOCAL_CACHE - Counter::get_instance().record("buffer_pool_handle_release_call", local_cache.size()); - for (const auto& pair : local_cache) { - pool.lp_map_.release_block(pair.first); - } + Counter::get_instance().record("buffer_pool_handle_release_call", + local_cache.size()); + for (const auto &pair : local_cache) { + pool.lp_map_.release_block(pair.first); + } #else - for (block_id_t block_id : local_cache) { - pool.lp_map_.release_block(block_id); - } + for (block_id_t block_id : local_cache) { + pool.lp_map_.release_block(block_id); + } #endif - local_cache.clear(); + local_cache.clear(); } \ No newline at end of file diff --git a/src/include/zvec/ailego/buffer/concurrentqueue.h b/src/include/zvec/ailego/buffer/concurrentqueue.h new file mode 100644 index 00000000..db4835b1 --- /dev/null +++ b/src/include/zvec/ailego/buffer/concurrentqueue.h @@ -0,0 +1,3747 @@ +// 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) 2013-2020, 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. + +// Also dual-licensed under the Boost Software License (see LICENSE.md) + +#pragma once + +#if defined(__GNUC__) && !defined(__INTEL_COMPILER) +// 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(_MSC_VER) && (!defined(_HAS_CXX17) || !_HAS_CXX17) +// VS2019 with /W4 warns about constant conditional expressions but unless /std=c++17 or higher +// does not support `if constexpr`, so we have no choice but to simply disable the warning +#pragma warning(push) +#pragma warning(disable: 4127) // conditional expression is constant +#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 +#include // used for thread exit synchronization + +// 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) || defined(__MVS__) || defined(MOODYCAMEL_NO_THREAD_LOCAL) +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. + inline thread_id_t thread_id() { static MOODYCAMEL_THREADLOCAL int x; return reinterpret_cast(&x); } +} } +#endif + +// Constexpr if +#ifndef MOODYCAMEL_CONSTEXPR_IF +#if (defined(_MSC_VER) && defined(_HAS_CXX17) && _HAS_CXX17) || __cplusplus > 201402L +#define MOODYCAMEL_CONSTEXPR_IF if constexpr +#define MOODYCAMEL_MAYBE_UNUSED [[maybe_unused]] +#else +#define MOODYCAMEL_CONSTEXPR_IF if +#define MOODYCAMEL_MAYBE_UNUSED +#endif +#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 MOODYCAMEL_CONSTEXPR_IF (true) +#define MOODYCAMEL_CATCH(...) else MOODYCAMEL_CONSTEXPR_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__) && !defined(__MVS__) +// Assume `thread_local` is fully supported in all other C++11 compilers/platforms +#define MOODYCAMEL_CPP11_THREAD_LOCAL_SUPPORTED // tentatively enabled for now; years ago 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 + +namespace moodycamel { namespace details { +#ifndef MOODYCAMEL_ALIGNAS +// VS2013 doesn't support alignas or alignof, and align() requires a constant literal +#if defined(_MSC_VER) && _MSC_VER <= 1800 +#define MOODYCAMEL_ALIGNAS(alignment) __declspec(align(alignment)) +#define MOODYCAMEL_ALIGNOF(obj) __alignof(obj) +#define MOODYCAMEL_ALIGNED_TYPE_LIKE(T, obj) typename details::Vs2013Aligned::value, T>::type + template struct Vs2013Aligned { }; // default, unsupported alignment + template struct Vs2013Aligned<1, T> { typedef __declspec(align(1)) T type; }; + template struct Vs2013Aligned<2, T> { typedef __declspec(align(2)) T type; }; + template struct Vs2013Aligned<4, T> { typedef __declspec(align(4)) T type; }; + template struct Vs2013Aligned<8, T> { typedef __declspec(align(8)) T type; }; + template struct Vs2013Aligned<16, T> { typedef __declspec(align(16)) T type; }; + template struct Vs2013Aligned<32, T> { typedef __declspec(align(32)) T type; }; + template struct Vs2013Aligned<64, T> { typedef __declspec(align(64)) T type; }; + template struct Vs2013Aligned<128, T> { typedef __declspec(align(128)) T type; }; + template struct Vs2013Aligned<256, T> { typedef __declspec(align(256)) T type; }; +#else + template struct identity { typedef T type; }; +#define MOODYCAMEL_ALIGNAS(alignment) alignas(alignment) +#define MOODYCAMEL_ALIGNOF(obj) alignof(obj) +#define MOODYCAMEL_ALIGNED_TYPE_LIKE(T, obj) alignas(alignof(obj)) typename details::identity::type +#endif +#endif +} } + + +// TSAN can false report races in lock-free code. To enable TSAN to be used from projects that use this one, +// we can apply per-function compile-time suppression. +// See https://clang.llvm.org/docs/ThreadSanitizer.html#has-feature-thread-sanitizer +#define MOODYCAMEL_NO_TSAN +#if defined(__has_feature) + #if __has_feature(thread_sanitizer) + #undef MOODYCAMEL_NO_TSAN + #define MOODYCAMEL_NO_TSAN __attribute__((no_sanitize("thread"))) + #endif // TSAN +#endif // TSAN + +// 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; + + // The number of times to spin before sleeping when waiting on a semaphore. + // Recommended values are on the order of 1000-10000 unless the number of + // consumer threads exceeds the number of idle cores (in which case try 0-100). + // Only affects instances of the BlockingConcurrentQueue. + static const int MAX_SEMA_SPINS = 10000; + + // Whether to recycle dynamically-allocated blocks into an internal free list or + // not. If false, only pre-allocated blocks (controlled by the constructor + // arguments) will be recycled, and all others will be `free`d back to the heap. + // Note that blocks consumed by explicit producers are only freed on destruction + // of the queue (not following destruction of the token) regardless of this trait. + static const bool RECYCLE_ALLOCATED_BLOCKS = false; + + +#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) + { + 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))); + // Note: extra parens around rhs of operator<< is MSVC bug: https://developercommunity2.visualstudio.com/t/C4554-triggers-when-both-lhs-and-rhs-is/10034931 + // silencing the bug requires #pragma warning(disable: 4554) around the calling code and has no effect when done here. + } + + 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 = left.load(std::memory_order_relaxed); + left.store(right.load(std::memory_order_relaxed), std::memory_order_relaxed); + right.store(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 + class ThreadExitNotifier; + + struct ThreadExitListener + { + typedef void (*callback_t)(void*); + callback_t callback; + void* userData; + + ThreadExitListener* next; // reserved for use by the ThreadExitNotifier + ThreadExitNotifier* chain; // reserved for use by the ThreadExitNotifier + }; + + class ThreadExitNotifier + { + public: + static void subscribe(ThreadExitListener* listener) + { + auto& tlsInst = instance(); + std::lock_guard guard(mutex()); + listener->next = tlsInst.tail; + listener->chain = &tlsInst; + tlsInst.tail = listener; + } + + static void unsubscribe(ThreadExitListener* listener) + { + std::lock_guard guard(mutex()); + if (!listener->chain) { + return; // race with ~ThreadExitNotifier + } + auto& tlsInst = *listener->chain; + listener->chain = nullptr; + 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."); + std::lock_guard guard(mutex()); + for (auto ptr = tail; ptr != nullptr; ptr = ptr->next) { + ptr->chain = nullptr; + ptr->callback(ptr->userData); + } + } + + // Thread-local + static inline ThreadExitNotifier& instance() + { + static thread_local ThreadExitNotifier notifier; + return notifier; + } + + static inline std::mutex& mutex() + { + // Must be static because the ThreadExitNotifier could be destroyed while unsubscribe is called + static std::mutex mutex; + return mutex; + } + + 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 = 32 * 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 + MOODYCAMEL_CONSTEXPR_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) + { + MOODYCAMEL_CONSTEXPR_IF (INITIAL_IMPLICIT_PRODUCER_HASH_SIZE == 0) return false; + else 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) + { + MOODYCAMEL_CONSTEXPR_IF (INITIAL_IMPLICIT_PRODUCER_HASH_SIZE == 0) return false; + else 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) + { + MOODYCAMEL_CONSTEXPR_IF (INITIAL_IMPLICIT_PRODUCER_HASH_SIZE == 0) return false; + else 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) + { + MOODYCAMEL_CONSTEXPR_IF (INITIAL_IMPLICIT_PRODUCER_HASH_SIZE == 0) return false; + else 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) + { + MOODYCAMEL_CONSTEXPR_IF (INITIAL_IMPLICIT_PRODUCER_HASH_SIZE == 0) return false; + else 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) + { + MOODYCAMEL_CONSTEXPR_IF (INITIAL_IMPLICIT_PRODUCER_HASH_SIZE == 0) return false; + else 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 constexpr 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)) { + 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_acq_rel) == 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), dynamicallyAllocated(true) + { +#ifdef MCDBGQ_TRACKMEM + owner = nullptr; +#endif + } + + template + inline bool is_empty() const + { + MOODYCAMEL_CONSTEXPR_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(MOODYCAMEL_MAYBE_UNUSED index_t i) + { + MOODYCAMEL_CONSTEXPR_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_acq_rel); + 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(MOODYCAMEL_MAYBE_UNUSED index_t i, size_t count) + { + MOODYCAMEL_CONSTEXPR_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_acq_rel); + assert(prevVal + count <= BLOCK_SIZE); + return prevVal + count == BLOCK_SIZE; + } + } + + template + inline void set_all_empty() + { + MOODYCAMEL_CONSTEXPR_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() + { + MOODYCAMEL_CONSTEXPR_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: + static_assert(std::alignment_of::value <= sizeof(T), "The queue does not support types with an alignment greater than their size at this time"); + MOODYCAMEL_ALIGNED_TYPE_LIKE(char[sizeof(T) * BLOCK_SIZE], T) elements; + 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; + 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; + 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. + + MOODYCAMEL_CONSTEXPR_IF (allocMode == CannotAlloc) { + return false; + } + else if (!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; + } + + MOODYCAMEL_CONSTEXPR_IF (!MOODYCAMEL_NOEXCEPT_CTOR(T, U, new (static_cast(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); + + MOODYCAMEL_CONSTEXPR_IF (!MOODYCAMEL_NOEXCEPT_CTOR(T, U, new (static_cast(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) / static_cast::type>(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 MOODYCAMEL_NO_TSAN 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) { + MOODYCAMEL_CONSTEXPR_IF (allocMode == CannotAlloc) { + // Failed to allocate, undo changes (but keep injected blocks) + pr_blockIndexFront = originalBlockIndexFront; + pr_blockIndexSlotsUsed = originalBlockIndexSlotsUsed; + this->tailBlock = startBlock == nullptr ? firstAllocatedBlock : startBlock; + return false; + } + else if (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; + } + + MOODYCAMEL_CONSTEXPR_IF (MOODYCAMEL_NOEXCEPT_CTOR(T, decltype(*itemFirst), new (static_cast(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) { + index_t stopIndex = (currentTailIndex & ~static_cast(BLOCK_SIZE - 1)) + static_cast(BLOCK_SIZE); + if (details::circular_less_than(newTailIndex, stopIndex)) { + stopIndex = newTailIndex; + } + MOODYCAMEL_CONSTEXPR_IF (MOODYCAMEL_NOEXCEPT_CTOR(T, decltype(*itemFirst), new (static_cast(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(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; + } + + MOODYCAMEL_CONSTEXPR_IF (!MOODYCAMEL_NOEXCEPT_CTOR(T, decltype(*itemFirst), new (static_cast(nullptr)) T(details::deref_noexcept(itemFirst)))) { + if (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) / static_cast::type>(BLOCK_SIZE)); + auto indexIndex = (localBlockIndexHead + offset) & (localBlockIndex->size - 1); + + // Iterate the blocks and dequeue + auto index = firstIndex; + do { + auto firstIndexInBlock = index; + index_t 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(); + + MOODYCAMEL_CONSTEXPR_IF (!MOODYCAMEL_NOEXCEPT_CTOR(T, U, new (static_cast(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; + + MOODYCAMEL_CONSTEXPR_IF (!MOODYCAMEL_NOEXCEPT_CTOR(T, U, new (static_cast(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; + } + +#ifdef _MSC_VER +#pragma warning(push) +#pragma warning(disable: 4706) // assignment within conditional expression +#endif + 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) { + index_t stopIndex = (currentTailIndex & ~static_cast(BLOCK_SIZE - 1)) + static_cast(BLOCK_SIZE); + if (details::circular_less_than(newTailIndex, stopIndex)) { + stopIndex = newTailIndex; + } + MOODYCAMEL_CONSTEXPR_IF (MOODYCAMEL_NOEXCEPT_CTOR(T, decltype(*itemFirst), new (static_cast(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(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; + } +#ifdef _MSC_VER +#pragma warning(pop) +#endif + + 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; + index_t 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 + } + size_t 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! + MOODYCAMEL_CONSTEXPR_IF (allocMode == CannotAlloc) { + return false; + } + else if (!new_block_index()) { + return false; + } + else { + 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) / static_cast::type>(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 + if (!Traits::RECYCLE_ALLOCATED_BLOCKS && block->dynamicallyAllocated) { + destroy(block); + } + else { + 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; + } + + MOODYCAMEL_CONSTEXPR_IF (canAlloc == CanAlloc) { + return create(); + } + else { + 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) + { +#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 + return ptr; + } + } + } + + 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() + { + MOODYCAMEL_CONSTEXPR_IF (INITIAL_IMPLICIT_PRODUCER_HASH_SIZE == 0) { + return; + } + else { + 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) + { + MOODYCAMEL_CONSTEXPR_IF (INITIAL_IMPLICIT_PRODUCER_HASH_SIZE == 0) { + return; + } + else { + // 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); + assert(mainHash != nullptr); // silence clang-tidy and MSVC warnings (hash cannot be null) + 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 - 1u; + + 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 - 1u; + auto empty = details::invalid_thread_id; +#ifdef MOODYCAMEL_CPP11_THREAD_LOCAL_SUPPORTED + auto reusable = details::invalid_thread_id2; + if (mainHash->entries[index].key.compare_exchange_strong(empty, id, std::memory_order_seq_cst, std::memory_order_relaxed) || + mainHash->entries[index].key.compare_exchange_strong(reusable, id, std::memory_order_seq_cst, std::memory_order_relaxed)) { +#else + if (mainHash->entries[index].key.compare_exchange_strong(empty, id, std::memory_order_seq_cst, 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)) { + size_t 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 = static_cast(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)) { + auto producer = static_cast(recycle_or_create_producer(false)); + if (producer == nullptr) { + implicitProducerHashCount.fetch_sub(1, std::memory_order_relaxed); + return nullptr; + } + +#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 - 1u; + auto empty = details::invalid_thread_id; +#ifdef MOODYCAMEL_CPP11_THREAD_LOCAL_SUPPORTED + auto reusable = details::invalid_thread_id2; + if (mainHash->entries[index].key.compare_exchange_strong(reusable, id, std::memory_order_seq_cst, std::memory_order_relaxed)) { + implicitProducerHashCount.fetch_sub(1, std::memory_order_relaxed); // already counted as a used slot + mainHash->entries[index].value = producer; + break; + } +#endif + if (mainHash->entries[index].key.compare_exchange_strong(empty, id, std::memory_order_seq_cst, std::memory_order_relaxed)) { + 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 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 - 1u; + probedKey = id; + if (hash->entries[index].key.compare_exchange_strong(probedKey, details::invalid_thread_id2, std::memory_order_seq_cst, std::memory_order_relaxed)) { + 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 void* aligned_malloc(size_t size) + { + MOODYCAMEL_CONSTEXPR_IF (std::alignment_of::value <= std::alignment_of::value) + return (Traits::malloc)(size); + else { + size_t alignment = std::alignment_of::value; + void* raw = (Traits::malloc)(size + alignment - 1 + sizeof(void*)); + if (!raw) + return nullptr; + char* ptr = details::align_for(reinterpret_cast(raw) + sizeof(void*)); + *(reinterpret_cast(ptr) - 1) = raw; + return ptr; + } + } + + template + static inline void aligned_free(void* ptr) + { + MOODYCAMEL_CONSTEXPR_IF (std::alignment_of::value <= std::alignment_of::value) + return (Traits::free)(ptr); + else + (Traits::free)(ptr ? *(reinterpret_cast(ptr) - 1) : nullptr); + } + + template + static inline U* create_array(size_t count) + { + assert(count > 0); + U* p = static_cast(aligned_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(); + } + aligned_free(p); + } + + template + static inline U* create() + { + void* p = aligned_malloc(sizeof(U)); + return p != nullptr ? new (p) U : nullptr; + } + + template + static inline U* create(A1&& a1) + { + void* p = aligned_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(); + aligned_free(p); + } + +private: + std::atomic producerListTail; + std::atomic producerCount; + + std::atomic initialBlockPoolIndex; + Block* initialBlockPool; + size_t initialBlockPoolSize; + +#ifndef 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 = static_cast(-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 = static_cast(-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(_MSC_VER) && (!defined(_HAS_CXX17) || !_HAS_CXX17) +#pragma warning(pop) +#endif + +#if defined(__GNUC__) && !defined(__INTEL_COMPILER) +#pragma GCC diagnostic pop +#endif \ No newline at end of file diff --git a/tests/core/algorithm/flat/flat_streamer_buffer_test.cpp b/tests/core/algorithm/flat/flat_streamer_buffer_test.cc similarity index 100% rename from tests/core/algorithm/flat/flat_streamer_buffer_test.cpp rename to tests/core/algorithm/flat/flat_streamer_buffer_test.cc diff --git a/tests/core/algorithm/flat/flat_streamer_buffer_time_test.cpp b/tests/core/algorithm/flat/flat_streamer_buffer_time_test.cc similarity index 100% rename from tests/core/algorithm/flat/flat_streamer_buffer_time_test.cpp rename to tests/core/algorithm/flat/flat_streamer_buffer_time_test.cc From 03e4dbce5437ed29f3db5bc79a58186541a1935b Mon Sep 17 00:00:00 2001 From: "yinzefeng.yzf" Date: Fri, 6 Feb 2026 21:08:24 +0800 Subject: [PATCH 03/11] modify buffer pool --- src/ailego/buffer/buffer_pool.cc | 257 ++++++++++++++ src/core/algorithm/hnsw/hnsw_entity.h | 4 +- .../algorithm/hnsw/hnsw_streamer_entity.cc | 2 +- src/core/utility/buffer1_storage.cc | 14 +- src/include/zvec/ailego/buffer/buffer_pool.h | 331 +++--------------- .../zvec/core/framework/index_storage.h | 57 ++- .../flat/flat_streamer_buffer_test.cc | 176 +++++----- .../flat/flat_streamer_buffer_time_test.cc | 129 ++++++- 8 files changed, 554 insertions(+), 416 deletions(-) create mode 100644 src/ailego/buffer/buffer_pool.cc diff --git a/src/ailego/buffer/buffer_pool.cc b/src/ailego/buffer/buffer_pool.cc new file mode 100644 index 00000000..061ead37 --- /dev/null +++ b/src/ailego/buffer/buffer_pool.cc @@ -0,0 +1,257 @@ +#include + +namespace zvec { +namespace ailego { + +void Counter::record(const std::string &name, int64_t value) { + auto it = static_counters.find(name); + if (it == static_counters.end()) { + auto counter = std::make_unique>(0); + it = static_counters.emplace(name, std::move(counter)).first; + } + it->second->fetch_add(value); +} + +void Counter::display() { + for (const auto &pair : static_counters) { + std::cout << pair.first << ": " << pair.second->load() << std::endl; + } +} + +int LRUCache::init(size_t block_size) { + block_size_ = block_size; + for(size_t i = 0; i < CATCH_QUEUE_NUM; i++) { + queues_.push_back(ConcurrentQueue(block_size)); + } + return 0; +} + +bool LRUCache::evict_single_block(BlockType &item) { + // std::cerr << "dequeue: " << item.first << std::endl; + bool found = false; + for(size_t i = 0; i < CATCH_QUEUE_NUM; i++) { + found = queues_[i].try_dequeue(item); + // std::cerr << "dequeue: " << found << std::endl; + if(found) { + break; + } + } + return found; +} + +bool LRUCache::add_single_block(const LPMap *lp_map, const BlockType &block, int block_type) { + bool ok = queues_[block_type].try_enqueue(block); + if(++evict_queue_insertions_ % block_size_ == 0) { + this->clear_dead_node(lp_map); + } + return ok; +} + +void LRUCache::clear_dead_node(const LPMap *lp_map) { + for(int i = 0; i < CATCH_QUEUE_NUM; i++) { + int clear_count = 0; + ConcurrentQueue tmp(block_size_); + BlockType item; + while(queues_[i].try_dequeue(item) && (clear_count++ < block_size_)) { + if(!lp_map->isDeadBlock(item)) { + tmp.try_enqueue(item); + } + } + while(tmp.try_dequeue(item)) { + if(!lp_map->isDeadBlock(item)) { + queues_[i].try_enqueue(item); + } + } + } +} + +void LPMap::init(size_t entry_num) { + if (entries_) { + delete[] entries_; + } + entry_num_ = entry_num; + entries_ = new Entry[entry_num_]; + for (size_t i = 0; i < entry_num_; i++) { + entries_[i].ref_count.store(std::numeric_limits::min()); + entries_[i].load_count.store(0); + entries_[i].buffer = nullptr; + } + cache_.init(entry_num); +} + +char* LPMap::acquire_block(block_id_t block_id) { + assert(block_id < entry_num_); + Entry &entry = entries_[block_id]; + if (entry.ref_count.load() == 0) { + ++entry.load_count; + // std::cout << entry.load_count.load() << std::endl; + } + ++entry.ref_count; + // std::cout << entry.ref_count.load() << std::endl; + if (entry.ref_count.load() < 0) { + // std::cout << "acquire block failed: " << block_id << ", " << entry.ref_count.load() << std::endl; + return nullptr; + } + return entry.buffer; +} + +void LPMap::release_block(block_id_t block_id) { + assert(block_id < entry_num_); + Entry &entry = entries_[block_id]; + int rc = entry.ref_count.fetch_sub(1); + // std::cout << "release block: " << block_id << ", " << entry.ref_count.load() << std::endl; + // assert(rc > 0); + if(entry.ref_count.load() == 0) { + LRUCache::BlockType block; + block.first = block_id; + block.second = entry.load_count.load(); + cache_.add_single_block(this, block, 0); + } +} + +char* LPMap::evict_block(block_id_t block_id) { + // std::cout << "evict block: " << block_id << std::endl; + assert(block_id < entry_num_); + Entry &entry = entries_[block_id]; + int expected = 0; + if (entry.ref_count.compare_exchange_strong( + expected, std::numeric_limits::min())) { + char *buffer = entry.buffer; + entry.buffer = nullptr; + return buffer; + } else { + return nullptr; + } +} + +char* LPMap::set_block_acquired(block_id_t block_id, char *buffer) { + assert(block_id < entry_num_); + Entry &entry = entries_[block_id]; + if (entry.ref_count.load() >= 0) { + entry.ref_count.fetch_add(1); + // std::cout << "Set block2 " << block_id << std::endl; + return entry.buffer; + } + // if (buffer == nullptr) std::cout << "Set block " << block_id << std::endl; + entry.buffer = buffer; + entry.ref_count.store(1); + entry.load_count.fetch_add(1); + return buffer; +} + +void LPMap::recycle(moodycamel::ConcurrentQueue &free_buffers) { + LRUCache::BlockType block; + do { + bool ok = cache_.evict_single_block(block); + if(!ok) { + return; + } + } while(isDeadBlock(block)); + // std::cout << "evict_block done: " << block.first << ", " << block.second << std::endl; + char *buffer = evict_block(block.first); + if (buffer) { + free_buffers.try_enqueue(buffer); + } +} + +VecBufferPool::VecBufferPool(const std::string &filename, size_t pool_capacity, size_t block_size) + : pool_capacity_(pool_capacity) { + fd_ = open(filename.c_str(), O_RDONLY); + if (fd_ < 0) { + throw std::runtime_error("Failed to open file: " + filename); + } + struct stat st; + if (fstat(fd_, &st) < 0) { + throw std::runtime_error("Failed to stat file: " + filename); + } + file_size_ = st.st_size; + + size_t buffer_num = pool_capacity_ / block_size; + size_t block_num = file_size_ / block_size + 500; + lp_map_.init(block_num); + for (size_t i = 0; i < buffer_num; i++) { + char *buffer = (char *)aligned_alloc(64, block_size); + if (buffer != nullptr) { + bool ok = free_buffers_.try_enqueue(buffer); + // if(!ok) std::cerr << i << std::endl; + } + } + std::cout << "buffer_num: " << buffer_num << std::endl; + std::cout << "entry_num: " << lp_map_.entry_num() << std::endl; +} + +VecBufferPoolHandle VecBufferPool::get_handle() { + return VecBufferPoolHandle(*this); +} + +char* VecBufferPool::acquire_buffer(block_id_t block_id, size_t offset, size_t size, int retry) { + char *buffer = lp_map_.acquire_block(block_id); + if (buffer) { + return buffer; + } + { + // std::cerr << "block_id: " << block_id << ", offset: " << offset << ", size: " << size << std::endl; + // std::lock_guard lock(mutex_); + bool found = free_buffers_.try_dequeue(buffer); + // std::cerr << "dequeue: " << found << std::endl; + if (!found) { + for (int i = 0; i < retry; i++) { + lp_map_.recycle(free_buffers_); + found = free_buffers_.try_dequeue(buffer); + // std::cerr << "dequeue: " << i << std::endl; + if (found) { + break; + } + } + } + if (!found) { + std::cerr << "Failed to get free buffer " << std::endl; + return nullptr; + } + } + + ssize_t read_bytes = pread(fd_, buffer, size, offset); + if (read_bytes != static_cast(size)) { + std::cerr << "Failed to read file at offset " << offset << std::endl; + exit(-1); + } + char *placed_buffer = nullptr; + { + std::lock_guard lock(mutex_); + placed_buffer = lp_map_.set_block_acquired(block_id, buffer); + } + if (placed_buffer != buffer) { + // another thread has set the block + free_buffers_.try_enqueue(buffer); + } + return placed_buffer; +} + +int VecBufferPool::get_meta(size_t offset, size_t length, char *buffer) { + ssize_t read_bytes = pread(fd_, buffer, length, offset); + if (read_bytes != static_cast(length)) { + std::cerr << "Failed to read file at offset " << offset << std::endl; + exit(-1); + } + return 0; +} + +char* VecBufferPoolHandle::get_block(size_t offset, size_t size, size_t block_id) { + char *buffer = pool.acquire_buffer(block_id, offset, size, 5); + return buffer; +} + +int VecBufferPoolHandle::get_meta(size_t offset, size_t length, char *buffer) { + return pool.get_meta(offset, length, buffer); +} + +void VecBufferPoolHandle::release_one(block_id_t block_id) { + pool.lp_map_.release_block(block_id); +} + +void VecBufferPoolHandle::acquire_one(block_id_t block_id) { + pool.lp_map_.acquire_block(block_id); +} + +} // namespace ailego +} // namespace zvec \ No newline at end of file diff --git a/src/core/algorithm/hnsw/hnsw_entity.h b/src/core/algorithm/hnsw/hnsw_entity.h index 363a7252..65fdae9e 100644 --- a/src/core/algorithm/hnsw/hnsw_entity.h +++ b/src/core/algorithm/hnsw/hnsw_entity.h @@ -147,8 +147,8 @@ struct Neighbors { Neighbors(uint32_t cnt_in, const node_id_t *data_in) : cnt{cnt_in}, data{data_in} {} - Neighbors(IndexStorage::MemoryBlock &&mem_block) - : neighbor_block{std::move(mem_block)} { + Neighbors(IndexStorage::MemoryBlock &mem_block) + : neighbor_block{mem_block} { auto hd = reinterpret_cast(neighbor_block.data()); cnt = hd->neighbor_cnt; data = hd->neighbors; diff --git a/src/core/algorithm/hnsw/hnsw_streamer_entity.cc b/src/core/algorithm/hnsw/hnsw_streamer_entity.cc index feafa573..734f11f1 100644 --- a/src/core/algorithm/hnsw/hnsw_streamer_entity.cc +++ b/src/core/algorithm/hnsw/hnsw_streamer_entity.cc @@ -127,7 +127,7 @@ const Neighbors HnswStreamerEntity::get_neighbors(level_t level, LOG_ERROR("Read neighbor header failed, ret=%zu", size); return Neighbors(); } - return Neighbors(std::move(neighbor_block)); + return Neighbors(neighbor_block); } //! Get vector data by key diff --git a/src/core/utility/buffer1_storage.cc b/src/core/utility/buffer1_storage.cc index 0ea591d9..1c582198 100644 --- a/src/core/utility/buffer1_storage.cc +++ b/src/core/utility/buffer1_storage.cc @@ -85,7 +85,6 @@ class Buffer1Storage : public IndexStorage { //! Read data from segment size_t read(size_t offset, const void **data, size_t len) override { - if (ailego_unlikely(offset + len > segment_->meta()->data_size)) { auto meta = segment_->meta(); if (offset > meta->data_size) { @@ -107,7 +106,8 @@ class Buffer1Storage : public IndexStorage { len = meta->data_size - offset; } size_t segment_offset = segment_->meta()->data_index + owner_->get_context_offset(); - data.reset(owner_->get_buffer(segment_offset, capacity_, segment_id_) + offset); + data.reset(owner_->buffer_pool_handle_.get(), segment_id_, owner_->get_buffer(segment_offset, capacity_, segment_id_) + offset); + // data.reset(owner_->get_buffer(segment_offset, capacity_, segment_id_) + offset); if (data.data()) { return len; } else { @@ -138,8 +138,8 @@ class Buffer1Storage : public IndexStorage { private: IndexMapping::Segment *segment_{}; Buffer1Storage *owner_{nullptr}; - size_t capacity_{}; size_t segment_id_{}; + size_t capacity_{}; }; //! Destructor @@ -162,9 +162,9 @@ class Buffer1Storage : public IndexStorage { int open(const std::string &path, bool /*create*/) override { LOG_INFO("open buffer storage 1"); file_name_ = path; - buffer_pool_ = std::make_unique(path, 10u * 1024 * 1024 * 1024, 2490368 * 2); + buffer_pool_ = std::make_shared(path, 10u * 1024 * 1024 * 1024, 2490368 * 2); buffer_pool_handle_ = - std::make_unique(buffer_pool_->get_handle()); + std::make_shared(buffer_pool_->get_handle()); int ret = ParseToMapping(); LOG_ERROR("segment count: %lu, max_segment_size: %lu", segments_.size(), max_segment_size_); if(ret != 0) { @@ -428,8 +428,8 @@ class Buffer1Storage : public IndexStorage { size_t max_segment_size_{0}; std::unique_ptr segment_buffer_{nullptr}; - std::unique_ptr buffer_pool_{nullptr}; - std::unique_ptr buffer_pool_handle_{nullptr}; + ailego::VecBufferPool::Pointer buffer_pool_{nullptr}; + ailego::VecBufferPoolHandle::Pointer buffer_pool_handle_{nullptr}; }; INDEX_FACTORY_REGISTER_STORAGE_ALIAS(BufferStorage, Buffer1Storage); diff --git a/src/include/zvec/ailego/buffer/buffer_pool.h b/src/include/zvec/ailego/buffer/buffer_pool.h index d86cffec..34c69d51 100644 --- a/src/include/zvec/ailego/buffer/buffer_pool.h +++ b/src/include/zvec/ailego/buffer/buffer_pool.h @@ -16,48 +16,41 @@ #include #include #include +#include #include "concurrentqueue.h" -using block_id_t = int; -using version_t = int; +namespace zvec { +namespace ailego { + +using block_id_t = size_t; +using version_t = size_t; + +class LPMap; class LRUCache { public: typedef std::pair BlockType; typedef moodycamel::ConcurrentQueue ConcurrentQueue; - int init(size_t block_size) { - for(int i = 0; i < CATCH_QUEUE_NUM; i++) { - queues_.push_back(ConcurrentQueue(block_size)); - } - return 0; - } - - BlockType evict_single_block() { - BlockType item; - for(int i = 0; i < CATCH_QUEUE_NUM; i++) { - bool found = queues_[i].try_dequeue(item); - if(found) { - break; - } - } - return item; - } - - bool add_single_block(const BlockType &block, int block_type) { - std::cout << "in LRU: " << block.first << ", " << block.second << std::endl; - return queues_[block_type].try_enqueue(block); - } + int init(size_t block_size); + + bool evict_single_block(BlockType &item); + + bool add_single_block(const LPMap *lp_map, const BlockType &block, int block_type); + + void clear_dead_node(const LPMap *lp_map); private: constexpr static size_t CATCH_QUEUE_NUM = 3; + int block_size_; std::vector queues_; + alignas(64) std::atomic evict_queue_insertions_{0}; }; class LPMap { struct Entry { - std::atomic ref_count; - std::atomic load_count; + alignas(64) std::atomic ref_count; + alignas(64) std::atomic load_count; char *buffer; }; @@ -67,206 +60,52 @@ class LPMap { delete[] entries_; } - void init(size_t entry_num) { - if (entries_) { - delete[] entries_; - } - entry_num_ = entry_num; - entries_ = new Entry[entry_num_]; - for (size_t i = 0; i < entry_num_; i++) { - entries_[i].ref_count.store(std::numeric_limits::min()); - entries_[i].load_count.store(0); - entries_[i].buffer = nullptr; - } - } + void init(size_t entry_num); - char *acquire_block(block_id_t block_id) { - assert(block_id < entry_num_); - Entry &entry = entries_[block_id]; - int rc = entry.ref_count.fetch_add(1); - if (rc < 0) { - return nullptr; - } - return entry.buffer; - } + char *acquire_block(block_id_t block_id); - void release_block(block_id_t block_id) { - assert(block_id < entry_num_); - Entry &entry = entries_[block_id]; - int rc = entry.ref_count.fetch_sub(1); - assert(rc >= 0); - if(rc == 0) { - LRUCache::BlockType block; - block.first = block_id; - block.second = entry.load_count.load(); - cache_.add_single_block(block, 0); - } - } + void release_block(block_id_t block_id); // need be called under lock - char *evict_block(block_id_t block_id) { - assert(block_id < entry_num_); - Entry &entry = entries_[block_id]; - int expected = 0; - if (entry.ref_count.compare_exchange_strong( - expected, std::numeric_limits::min())) { - char *buffer = entry.buffer; - entry.buffer = nullptr; - return buffer; - } else { - return nullptr; - } - } + char *evict_block(block_id_t block_id); // need be called under lock - char *set_block_acquired(block_id_t block_id, char *buffer) { - // std::cout << "Set block " << block_id << std::endl; - assert(block_id < entry_num_); - Entry &entry = entries_[block_id]; - if (entry.ref_count.load() >= 0) { - entry.ref_count.fetch_add(1); - return entry.buffer; - } - entry.buffer = buffer; - entry.ref_count.store(1); - entry.load_count.fetch_add(1); - return buffer; - } + char *set_block_acquired(block_id_t block_id, char *buffer); // need be called under lock - void recycle(std::queue &free_buffers) { - LRUCache::BlockType block; - do { - block = cache_.evict_single_block(); - } while(isDeadBlock(block)); - char *buffer = evict_block(block.first); - if (buffer) { - free_buffers.push(buffer); - } - } + void recycle(moodycamel::ConcurrentQueue &free_buffers); size_t entry_num() const { return entry_num_; } - private: - Entry *entries_; - size_t entry_num_; - LRUCache cache_; - - bool isDeadBlock(LRUCache::BlockType block) { + bool isDeadBlock(LRUCache::BlockType block) const { Entry &entry = entries_[block.first]; - return block.second == entry.load_count.load(); - } -}; - -class VecBufferPool; - -struct VecBufferPoolHandle { - VecBufferPoolHandle(VecBufferPool &pool); - VecBufferPoolHandle(VecBufferPoolHandle &&other) - : pool(other.pool), - local_cache(std::move(other.local_cache)), - hit_num_(other.hit_num_) { - other.local_cache.clear(); - other.hit_num_ = 0; + return block.second != entry.load_count.load(); } - ~VecBufferPoolHandle(); - char *get_block(size_t offset, size_t size, size_t block_id); - - int get_meta(size_t offset, size_t length, char *buffer); - - void release_all(); - - VecBufferPool &pool; -#ifdef USE_LOCAL_CACHE - // std::unordered_map local_cache; - phmap::flat_hash_map local_cache; -#else - std::vector local_cache; -#endif - int hit_num_; + private: + size_t entry_num_{0}; + Entry *entries_{nullptr}; + LRUCache cache_; }; +class VecBufferPoolHandle; + class VecBufferPool { public: - VecBufferPool(const std::string &filename, size_t pool_capacity, size_t block_size) - : pool_capacity_(pool_capacity) { - fd_ = open(filename.c_str(), O_RDONLY); - if (fd_ < 0) { - throw std::runtime_error("Failed to open file: " + filename); - } - struct stat st; - if (fstat(fd_, &st) < 0) { - throw std::runtime_error("Failed to stat file: " + filename); - } - file_size_ = st.st_size; - - size_t buffer_num = pool_capacity_ / block_size; - lp_map_.init(buffer_num); - for (size_t i = 0; i < buffer_num; i++) { - char *buffer = (char *)aligned_alloc(64, block_size); - free_buffers_.push(buffer); - } - std::cout << "buffer_num: " << buffer_num << std::endl; - std::cout << "entry_num: " << lp_map_.entry_num() << std::endl; - } + typedef std::shared_ptr Pointer; + + VecBufferPool(const std::string &filename, size_t pool_capacity, size_t block_size); ~VecBufferPool() { close(fd_); } - VecBufferPoolHandle get_handle() { - return VecBufferPoolHandle(*this); - } + VecBufferPoolHandle get_handle(); - char *acquire_buffer(block_id_t block_id, size_t offset, size_t size, int retry = 0) { - char *buffer = lp_map_.acquire_block(block_id); - if (buffer) { - return buffer; - } - { - std::lock_guard lock(mutex_); - if (free_buffers_.empty()) { - for (int i = 0; i < retry; i++) { - lp_map_.recycle(free_buffers_); - if (!free_buffers_.empty()) { - break; - } - } - } - if (free_buffers_.empty()) { - return nullptr; - } - buffer = free_buffers_.front(); - free_buffers_.pop(); - } - - ssize_t read_bytes = pread(fd_, buffer, size, offset); - if (read_bytes != static_cast(size)) { - std::cerr << "Failed to read file at offset " << offset << std::endl; - exit(-1); - } - - { - std::lock_guard lock(mutex_); - char *placed_buffer = lp_map_.set_block_acquired(block_id, buffer); - if (placed_buffer != buffer) { - // another thread has set the block - free_buffers_.push(buffer); - } - return placed_buffer; - } - } + char *acquire_buffer(block_id_t block_id, size_t offset, size_t size, int retry = 0); - int get_meta(size_t offset, size_t length, char *buffer) { - ssize_t read_bytes = pread(fd_, buffer, length, offset); - if (read_bytes != static_cast(length)) { - std::cerr << "Failed to read file at offset " << offset << std::endl; - exit(-1); - } - return 0; - } + int get_meta(size_t offset, size_t length, char *buffer); size_t file_size() const { return file_size_; @@ -282,86 +121,32 @@ class VecBufferPool { private: std::mutex mutex_; - std::queue free_buffers_; + moodycamel::ConcurrentQueue free_buffers_; }; - -struct Counter { - ~Counter() = default; - - static Counter &get_instance() { - static Counter instance; - return instance; - } - - void record(const std::string &name, int64_t value) { - auto it = static_counters.find(name); - if (it == static_counters.end()) { - auto counter = std::make_unique>(0); - it = static_counters.emplace(name, std::move(counter)).first; - } - it->second->fetch_add(value); +struct VecBufferPoolHandle { + VecBufferPoolHandle(VecBufferPool &pool) : pool(pool), hit_num_(0) {}; + VecBufferPoolHandle(VecBufferPoolHandle &&other) + : pool(other.pool), + hit_num_(other.hit_num_) { + other.hit_num_ = 0; } + + ~VecBufferPoolHandle() = default; - void display() { - for (const auto &pair : static_counters) { - std::cout << pair.first << ": " << pair.second->load() << std::endl; - } - } + typedef std::shared_ptr Pointer; - void clear() { - static_counters.clear(); - } + char *get_block(size_t offset, size_t size, size_t block_id); - private: - Counter() {} - std::map>> static_counters; -}; + int get_meta(size_t offset, size_t length, char *buffer); -VecBufferPoolHandle::VecBufferPoolHandle(VecBufferPool &pool) - : pool(pool), hit_num_(0) {} -VecBufferPoolHandle::~VecBufferPoolHandle() { - Counter::get_instance().record("buffer_pool_handle_hit_num", hit_num_); - release_all(); -} - -char *VecBufferPoolHandle::get_block(size_t offset, size_t size, size_t block_id) { -#ifdef USE_LOCAL_CACHE - auto it = local_cache.find(block_id); - if (it != local_cache.end()) { - hit_num_++; - return it->second; - } -#endif - - char *buffer = pool.acquire_buffer(block_id, offset, size, 3); - if (buffer) { -#ifdef USE_LOCAL_CACHE - local_cache[block_id] = buffer; -#else - local_cache.push_back(block_id); -#endif - return buffer; - } + void release_one(block_id_t block_id); - return nullptr; -} + void acquire_one(block_id_t block_id); -int VecBufferPoolHandle::get_meta(size_t offset, size_t length, char *out) { - return pool.get_meta(offset, length, out); -} + VecBufferPool &pool; + int hit_num_; +}; -void VecBufferPoolHandle::release_all() { -#ifdef USE_LOCAL_CACHE - Counter::get_instance().record("buffer_pool_handle_release_call", - local_cache.size()); - for (const auto &pair : local_cache) { - pool.lp_map_.release_block(pair.first); - } -#else - for (block_id_t block_id : local_cache) { - pool.lp_map_.release_block(block_id); - } -#endif - local_cache.clear(); -} \ No newline at end of file +} // namespace ailego +} // namespace zvec \ No newline at end of file diff --git a/src/include/zvec/core/framework/index_storage.h b/src/include/zvec/core/framework/index_storage.h index 8673d63e..346b8da4 100644 --- a/src/include/zvec/core/framework/index_storage.h +++ b/src/include/zvec/core/framework/index_storage.h @@ -14,7 +14,7 @@ #pragma once -#include +#include #include #include #include @@ -37,10 +37,11 @@ class IndexStorage : public IndexModule { }; MemoryBlock() {} - MemoryBlock(ailego::BufferHandle::Pointer &&buffer_handle) - : type_(MemoryBlockType::MBT_BUFFERPOOL), - buffer_handle_(std::move(buffer_handle)) { - data_ = buffer_handle_->pin_vector_data(); + MemoryBlock(ailego::VecBufferPoolHandle* buffer_pool_handle, int block_id, void *data) + : type_(MemoryBlockType::MBT_BUFFERPOOL) { + buffer_pool_handle_ = buffer_pool_handle; + buffer_block_id_ = block_id; + data_ = data; } MemoryBlock(void *data) : type_(MemoryBlockType::MBT_MMAP), data_(data) {} @@ -50,7 +51,8 @@ class IndexStorage : public IndexModule { this->reset(rhs.data_); break; case MemoryBlockType::MBT_BUFFERPOOL: - this->reset(rhs.buffer_handle_); + this->reset(rhs.buffer_pool_handle_, rhs.buffer_block_id_, rhs.data_); + buffer_pool_handle_->acquire_one(buffer_block_id_); break; default: break; @@ -63,7 +65,7 @@ class IndexStorage : public IndexModule { this->reset(std::move(rhs.data_)); break; case MemoryBlockType::MBT_BUFFERPOOL: - this->reset(std::move(rhs.buffer_handle_)); + this->reset(std::move(rhs.buffer_pool_handle_), std::move(rhs.buffer_block_id_), std::move(rhs.data_)); break; default: break; @@ -77,7 +79,8 @@ class IndexStorage : public IndexModule { this->reset(rhs.data_); break; case MemoryBlockType::MBT_BUFFERPOOL: - this->reset(rhs.buffer_handle_); + this->reset(rhs.buffer_pool_handle_, rhs.buffer_block_id_, rhs.data_); + buffer_pool_handle_->acquire_one(buffer_block_id_); break; default: break; @@ -93,7 +96,7 @@ class IndexStorage : public IndexModule { this->reset(std::move(rhs.data_)); break; case MemoryBlockType::MBT_BUFFERPOOL: - this->reset(std::move(rhs.buffer_handle_)); + this->reset(std::move(rhs.buffer_pool_handle_), std::move(rhs.buffer_block_id_), std::move(rhs.data_)); break; default: break; @@ -107,9 +110,8 @@ class IndexStorage : public IndexModule { case MemoryBlockType::MBT_MMAP: break; case MemoryBlockType::MBT_BUFFERPOOL: - if (buffer_handle_) { - buffer_handle_->unpin_vector_data(); - // buffer_handle_.reset(); + if (buffer_pool_handle_) { + buffer_pool_handle_->release_one(buffer_block_id_); } break; default: @@ -122,34 +124,20 @@ class IndexStorage : public IndexModule { return data_; } - void reset(ailego::BufferHandle::Pointer &buffer_handle) { + void reset(ailego::VecBufferPoolHandle* buffer_pool_handle, int block_id, void *data) { if (type_ == MemoryBlockType::MBT_BUFFERPOOL) { - buffer_handle_->unpin_vector_data(); - buffer_handle_.reset(); + buffer_pool_handle->release_one(buffer_block_id_); } type_ = MemoryBlockType::MBT_BUFFERPOOL; - if (buffer_handle) { - buffer_handle_.reset(buffer_handle.release()); - } - data_ = buffer_handle_->pin_vector_data(); - } - - void reset(ailego::BufferHandle::Pointer &&buffer_handle) { - if (type_ == MemoryBlockType::MBT_BUFFERPOOL) { - buffer_handle_->unpin_vector_data(); - buffer_handle_.reset(); - } - type_ = MemoryBlockType::MBT_BUFFERPOOL; - if (buffer_handle) { - buffer_handle_ = std::move(buffer_handle); - } - data_ = buffer_handle_->pin_vector_data(); + buffer_pool_handle_ = buffer_pool_handle; + buffer_block_id_ = block_id; + data_ = data; } void reset(void *data) { if (type_ == MemoryBlockType::MBT_BUFFERPOOL) { - buffer_handle_->unpin_vector_data(); - buffer_handle_.reset(); + buffer_pool_handle_->release_one(buffer_block_id_); + buffer_pool_handle_ = nullptr; } type_ = MemoryBlockType::MBT_MMAP; data_ = data; @@ -157,7 +145,8 @@ class IndexStorage : public IndexModule { MemoryBlockType type_{MBT_UNKNOWN}; void *data_{nullptr}; - mutable ailego::BufferHandle::Pointer buffer_handle_{nullptr}; + mutable ailego::VecBufferPoolHandle* buffer_pool_handle_; + int buffer_block_id_{0}; }; struct SegmentData { diff --git a/tests/core/algorithm/flat/flat_streamer_buffer_test.cc b/tests/core/algorithm/flat/flat_streamer_buffer_test.cc index 62b25e23..fbc404b4 100644 --- a/tests/core/algorithm/flat/flat_streamer_buffer_test.cc +++ b/tests/core/algorithm/flat/flat_streamer_buffer_test.cc @@ -50,7 +50,6 @@ void FlatStreamerTest::TearDown(void) { } TEST_F(FlatStreamerTest, TestLinearSearch) { - BufferManager::Instance().init(300 * 1024 / 2 * 1024, 1); IndexStreamer::Pointer write_streamer = IndexFactory::CreateStreamer("FlatStreamer"); ASSERT_TRUE(write_streamer != nullptr); @@ -165,31 +164,33 @@ TEST_F(FlatStreamerTest, TestLinearSearch) { ASSERT_EQ(i == cnt - 1 ? i - 1 : i + 1, result2[1].key()); ASSERT_EQ(i == 0 ? 2 : (i == cnt - 1 ? i - 2 : i - 1), result2[2].key()); } + cout << "Elapsed time: " << elapsed_time.milli_seconds() << " ms" << endl; read_streamer->close(); read_streamer.reset(); - cout << "Elapsed time: " << elapsed_time.milli_seconds() << " ms" << endl; } -TEST_F(FlatStreamerTest, TestLinearSearchMMap) { - BufferManager::Instance().init(3 * 1024 / 2 * 1024, 1); +TEST_F(FlatStreamerTest, TestLinearSearchWithLRU) { + constexpr size_t static dim = 1600; IndexStreamer::Pointer write_streamer = IndexFactory::CreateStreamer("FlatStreamer"); ASSERT_TRUE(write_streamer != nullptr); Params params; - ASSERT_EQ(0, write_streamer->init(*index_meta_ptr_, params)); + IndexMeta meta = IndexMeta(IndexMeta::DataType::DT_FP32, dim); + meta.set_metric("SquaredEuclidean", 0, Params()); + ASSERT_EQ(0, write_streamer->init(meta, params)); auto storage = IndexFactory::CreateStorage("MMapFileStorage"); ASSERT_NE(nullptr, storage); Params stg_params; ASSERT_EQ(0, storage->init(stg_params)); - ASSERT_EQ(0, storage->open(dir_ + "/Test/LinearSearchMMap", true)); + ASSERT_EQ(0, storage->open(dir_ + "/Test/LinearSearchWithLRU", true)); ASSERT_EQ(0, write_streamer->open(storage)); auto ctx = write_streamer->create_context(); ASSERT_TRUE(!!ctx); - size_t cnt = 10000UL; + size_t cnt = 1000000UL; IndexQueryMeta qmeta(IndexMeta::DT_FP32, dim); for (size_t i = 0; i < cnt; i++) { NumericalVector vec(dim); @@ -202,18 +203,19 @@ TEST_F(FlatStreamerTest, TestLinearSearchMMap) { write_streamer->close(); write_streamer.reset(); - ElapsedTime elapsed_time; + IndexStreamer::Pointer read_streamer = IndexFactory::CreateStreamer("FlatStreamer"); - ASSERT_EQ(0, read_streamer->init(*index_meta_ptr_, params)); - auto read_storage = IndexFactory::CreateStorage("MMapFileStorage"); + ASSERT_EQ(0, read_streamer->init(meta, params)); + auto read_storage = IndexFactory::CreateStorage("BufferStorage"); ASSERT_NE(nullptr, read_storage); ASSERT_EQ(0, read_storage->init(stg_params)); - ASSERT_EQ(0, read_storage->open(dir_ + "/Test/LinearSearchMMap", false)); + ASSERT_EQ(0, read_storage->open(dir_ + "/Test/LinearSearchWithLRU", false)); ASSERT_EQ(0, read_streamer->open(read_storage)); size_t topk = 3; auto provider = read_streamer->create_provider(); - for (size_t i = 0; i < cnt; i += 1) { + ElapsedTime elapsed_time; + for (size_t i = 0; i < 10; i += 1) { NumericalVector vec(dim); for (size_t j = 0; j < dim; ++j) { vec[j] = i; @@ -241,122 +243,132 @@ TEST_F(FlatStreamerTest, TestLinearSearchMMap) { ASSERT_EQ(i == cnt - 1 ? i - 1 : i + 1, result2[1].key()); ASSERT_EQ(i == 0 ? 2 : (i == cnt - 1 ? i - 2 : i - 1), result2[2].key()); } - - ctx->set_topk(100U); - NumericalVector vec(dim); - for (size_t j = 0; j < dim; ++j) { - vec[j] = 10.1f; - } - ASSERT_EQ(0, read_streamer->search_bf_impl(vec.data(), qmeta, ctx)); - auto &result = ctx->result(); - ASSERT_EQ(100U, result.size()); - ASSERT_EQ(10, result[0].key()); - ASSERT_EQ(11, result[1].key()); - ASSERT_EQ(5, result[10].key()); - ASSERT_EQ(0, result[20].key()); - ASSERT_EQ(30, result[30].key()); - ASSERT_EQ(35, result[35].key()); - ASSERT_EQ(99, result[99].key()); + cout << "Elapsed time: " << elapsed_time.milli_seconds() << " ms" << endl; read_streamer->close(); read_streamer.reset(); - cout << "Elapsed time: " << elapsed_time.milli_seconds() << " ms" << endl; } -TEST_F(FlatStreamerTest, TestBufferStorage) { - BufferManager::Instance().init(10 * 1024 * 1024, 1); - IndexStreamer::Pointer streamer = +TEST_F(FlatStreamerTest, TestLinearSearchMMap) { + IndexStreamer::Pointer write_streamer = IndexFactory::CreateStreamer("FlatStreamer"); - ASSERT_TRUE(streamer != nullptr); - const int dim = 16; - IndexMeta meta = IndexMeta(IndexMeta::DT_FP32, dim); - meta.set_metric("SquaredEuclidean", 0, Params()); + ASSERT_TRUE(write_streamer != nullptr); Params params; - EXPECT_EQ(0, streamer->init(meta, params)); + ASSERT_EQ(0, write_streamer->init(*index_meta_ptr_, params)); auto storage = IndexFactory::CreateStorage("MMapFileStorage"); ASSERT_NE(nullptr, storage); Params stg_params; - EXPECT_EQ(0, storage->init(stg_params)); - EXPECT_EQ(0, storage->open(dir_ + "/Test/LinearSearch", true)); - EXPECT_EQ(0, streamer->open(storage)); + ASSERT_EQ(0, storage->init(stg_params)); + ASSERT_EQ(0, storage->open(dir_ + "/Test/LinearSearchMMap", true)); + ASSERT_EQ(0, write_streamer->open(storage)); - auto ctx = streamer->create_context(); + auto ctx = write_streamer->create_context(); ASSERT_TRUE(!!ctx); - size_t cnt = 1000UL; + size_t cnt = 10000UL; IndexQueryMeta qmeta(IndexMeta::DT_FP32, dim); for (size_t i = 0; i < cnt; i++) { NumericalVector vec(dim); for (size_t j = 0; j < dim; ++j) { vec[j] = i; } - streamer->add_impl(i, vec.data(), qmeta, ctx); + write_streamer->add_impl(i, vec.data(), qmeta, ctx); } - streamer->flush(0UL); - streamer.reset(); + write_streamer->flush(0UL); + write_streamer->close(); + write_streamer.reset(); IndexStreamer::Pointer read_streamer = IndexFactory::CreateStreamer("FlatStreamer"); - ASSERT_TRUE(read_streamer != nullptr); - EXPECT_EQ(0, read_streamer->init(meta, params)); - auto read_storage = IndexFactory::CreateStorage("BufferStorage"); + ASSERT_EQ(0, read_streamer->init(*index_meta_ptr_, params)); + auto read_storage = IndexFactory::CreateStorage("MMapFileStorage"); ASSERT_NE(nullptr, read_storage); - EXPECT_EQ(0, read_storage->init(stg_params)); - EXPECT_EQ(0, read_storage->open(dir_ + "/Test/LinearSearch", false)); - EXPECT_EQ(0, read_streamer->open(read_storage)); - auto read_ctx = read_streamer->create_context(); - auto provider = read_streamer->create_provider(); - + ASSERT_EQ(0, read_storage->init(stg_params)); + ASSERT_EQ(0, read_storage->open(dir_ + "/Test/LinearSearchMMap", false)); + ASSERT_EQ(0, read_streamer->open(read_storage)); size_t topk = 3; + auto provider = read_streamer->create_provider(); for (size_t i = 0; i < cnt; i += 1) { NumericalVector vec(dim); for (size_t j = 0; j < dim; ++j) { vec[j] = i; } - read_ctx->set_topk(topk); - EXPECT_EQ(0, read_streamer->search_impl(vec.data(), qmeta, read_ctx)); - auto &result1 = read_ctx->result(); - EXPECT_EQ(topk, result1.size()); + ctx->set_topk(topk); + ASSERT_EQ(0, read_streamer->search_impl(vec.data(), qmeta, ctx)); + auto &result1 = ctx->result(); + ASSERT_EQ(topk, result1.size()); + IndexStorage::MemoryBlock block; + ASSERT_EQ(0, provider->get_vector(result1[0].key(), block)); + const float *data = (float *)block.data(); for (size_t j = 0; j < dim; ++j) { - const float *data = (float *)provider->get_vector(result1[0].key()); - EXPECT_EQ(data[j], i); + ASSERT_EQ(data[j], i); } - EXPECT_EQ(i, result1[0].key()); + ASSERT_EQ(i, result1[0].key()); for (size_t j = 0; j < dim; ++j) { vec[j] = i + 0.1f; } - read_ctx->set_topk(topk); - EXPECT_EQ(0, read_streamer->search_impl(vec.data(), qmeta, read_ctx)); - auto &result2 = read_ctx->result(); - EXPECT_EQ(topk, result2.size()); - EXPECT_EQ(i, result2[0].key()); - EXPECT_EQ(i == cnt - 1 ? i - 1 : i + 1, result2[1].key()); - EXPECT_EQ(i == 0 ? 2 : (i == cnt - 1 ? i - 2 : i - 1), result2[2].key()); + ctx->set_topk(topk); + ASSERT_EQ(0, read_streamer->search_impl(vec.data(), qmeta, ctx)); + auto &result2 = ctx->result(); + ASSERT_EQ(topk, result2.size()); + ASSERT_EQ(i, result2[0].key()); + ASSERT_EQ(i == cnt - 1 ? i - 1 : i + 1, result2[1].key()); + ASSERT_EQ(i == 0 ? 2 : (i == cnt - 1 ? i - 2 : i - 1), result2[2].key()); } - read_ctx->set_topk(100U); + ctx->set_topk(100U); NumericalVector vec(dim); for (size_t j = 0; j < dim; ++j) { vec[j] = 10.1f; } - EXPECT_EQ(0, read_streamer->search_bf_impl(vec.data(), qmeta, read_ctx)); - auto &result = read_ctx->result(); - EXPECT_EQ(100U, result.size()); - EXPECT_EQ(10, result[0].key()); - EXPECT_EQ(11, result[1].key()); - EXPECT_EQ(5, result[10].key()); - EXPECT_EQ(0, result[20].key()); - EXPECT_EQ(30, result[30].key()); - EXPECT_EQ(35, result[35].key()); - EXPECT_EQ(99, result[99].key()); - - read_streamer->flush(0UL); + ASSERT_EQ(0, read_streamer->search_bf_impl(vec.data(), qmeta, ctx)); + auto &result = ctx->result(); + ASSERT_EQ(100U, result.size()); + ASSERT_EQ(10, result[0].key()); + ASSERT_EQ(11, result[1].key()); + ASSERT_EQ(5, result[10].key()); + ASSERT_EQ(0, result[20].key()); + ASSERT_EQ(30, result[30].key()); + ASSERT_EQ(35, result[35].key()); + ASSERT_EQ(99, result[99].key()); + + ElapsedTime elapsed_time; + for (size_t i = 0; i < cnt; i += 1) { + NumericalVector vec(dim); + for (size_t j = 0; j < dim; ++j) { + vec[j] = i; + } + ctx->set_topk(topk); + ASSERT_EQ(0, read_streamer->search_impl(vec.data(), qmeta, ctx)); + auto &result1 = ctx->result(); + ASSERT_EQ(topk, result1.size()); + IndexStorage::MemoryBlock block; + ASSERT_EQ(0, provider->get_vector(result1[0].key(), block)); + const float *data = (float *)block.data(); + for (size_t j = 0; j < dim; ++j) { + ASSERT_EQ(data[j], i); + } + ASSERT_EQ(i, result1[0].key()); + + for (size_t j = 0; j < dim; ++j) { + vec[j] = i + 0.1f; + } + ctx->set_topk(topk); + ASSERT_EQ(0, read_streamer->search_impl(vec.data(), qmeta, ctx)); + auto &result2 = ctx->result(); + ASSERT_EQ(topk, result2.size()); + ASSERT_EQ(i, result2[0].key()); + ASSERT_EQ(i == cnt - 1 ? i - 1 : i + 1, result2[1].key()); + ASSERT_EQ(i == 0 ? 2 : (i == cnt - 1 ? i - 2 : i - 1), result2[2].key()); + } + + read_streamer->close(); read_streamer.reset(); + cout << "Elapsed time: " << elapsed_time.milli_seconds() << " ms" << endl; } - #if defined(__GNUC__) || defined(__GNUG__) #pragma GCC diagnostic pop #endif \ No newline at end of file diff --git a/tests/core/algorithm/flat/flat_streamer_buffer_time_test.cc b/tests/core/algorithm/flat/flat_streamer_buffer_time_test.cc index c919e9fe..435ecccc 100644 --- a/tests/core/algorithm/flat/flat_streamer_buffer_time_test.cc +++ b/tests/core/algorithm/flat/flat_streamer_buffer_time_test.cc @@ -83,7 +83,7 @@ TEST_F(FlatStreamerTest, TestLinearSearchMMap) { IndexStreamer::Pointer read_streamer = IndexFactory::CreateStreamer("FlatStreamer"); ASSERT_EQ(0, read_streamer->init(*index_meta_ptr_, params)); - auto read_storage = IndexFactory::CreateStorage("BufferStorage"); + auto read_storage = IndexFactory::CreateStorage("MMapFileStorage"); ASSERT_NE(nullptr, read_storage); ASSERT_EQ(0, read_storage->init(stg_params)); ASSERT_EQ(0, read_storage->open(dir_ + "/Test/LinearSearchMMap", false)); @@ -113,26 +113,121 @@ TEST_F(FlatStreamerTest, TestLinearSearchMMap) { // ASSERT_EQ(i == 0 ? 2 : (i == cnt - 1 ? i - 2 : i - 1), result2[2].key()); } cout << "Elapsed time: " << elapsed_time.micro_seconds() << " us" << endl; + for (size_t i = 0; i < cnt; i += 1) { + NumericalVector vec(dim); + for (size_t j = 0; j < dim; ++j) { + vec[j] = i; + } + ctx->set_topk(topk); + ASSERT_EQ(0, read_streamer->search_impl(vec.data(), qmeta, ctx)); + // auto &result1 = ctx->result(); + // ASSERT_EQ(topk, result1.size()); + // ASSERT_EQ(i, result1[0].key()); - // ctx->set_topk(100U); - // NumericalVector vec(dim); - // for (size_t j = 0; j < dim; ++j) { - // vec[j] = 10.1f; - // } - // ASSERT_EQ(0, read_streamer->search_bf_impl(vec.data(), qmeta, ctx)); - // auto &result = ctx->result(); - // ASSERT_EQ(100U, result.size()); - // ASSERT_EQ(10, result[0].key()); - // ASSERT_EQ(11, result[1].key()); - // ASSERT_EQ(5, result[10].key()); - // ASSERT_EQ(0, result[20].key()); - // ASSERT_EQ(30, result[30].key()); - // ASSERT_EQ(35, result[35].key()); - // ASSERT_EQ(99, result[99].key()); + // for (size_t j = 0; j < dim; ++j) { + // vec[j] = i + 0.1f; + // } + // ctx->set_topk(topk); + // ASSERT_EQ(0, read_streamer->search_impl(vec.data(), qmeta, ctx)); + // auto &result2 = ctx->result(); + // ASSERT_EQ(topk, result2.size()); + // ASSERT_EQ(i, result2[0].key()); + // ASSERT_EQ(i == cnt - 1 ? i - 1 : i + 1, result2[1].key()); + // ASSERT_EQ(i == 0 ? 2 : (i == cnt - 1 ? i - 2 : i - 1), result2[2].key()); + } + cout << "Elapsed time: " << elapsed_time.micro_seconds() << " us" << endl; + read_streamer->close(); + read_streamer.reset(); +} +TEST_F(FlatStreamerTest, TestLinearSearchBuffer) { + IndexStreamer::Pointer write_streamer = + IndexFactory::CreateStreamer("FlatStreamer"); + ASSERT_TRUE(write_streamer != nullptr); + + Params params; + ASSERT_EQ(0, write_streamer->init(*index_meta_ptr_, params)); + auto storage = IndexFactory::CreateStorage("MMapFileStorage"); + ASSERT_NE(nullptr, storage); + Params stg_params; + ASSERT_EQ(0, storage->init(stg_params)); + ASSERT_EQ(0, storage->open(dir_ + "/Test/LinearSearchBuffer", true)); + ASSERT_EQ(0, write_streamer->open(storage)); + + auto ctx = write_streamer->create_context(); + ASSERT_TRUE(!!ctx); + + size_t data_cnt = 300000UL, cnt = 500UL; + IndexQueryMeta qmeta(IndexMeta::DT_FP32, dim); + for (size_t i = 0; i < data_cnt; i++) { + NumericalVector vec(dim); + for (size_t j = 0; j < dim; ++j) { + vec[j] = i; + } + write_streamer->add_impl(i, vec.data(), qmeta, ctx); + } + write_streamer->flush(0UL); + write_streamer->close(); + write_streamer.reset(); + + IndexStreamer::Pointer read_streamer = + IndexFactory::CreateStreamer("FlatStreamer"); + ASSERT_EQ(0, read_streamer->init(*index_meta_ptr_, params)); + auto read_storage = IndexFactory::CreateStorage("BufferStorage"); + ASSERT_NE(nullptr, read_storage); + ASSERT_EQ(0, read_storage->init(stg_params)); + ASSERT_EQ(0, read_storage->open(dir_ + "/Test/LinearSearchBuffer", false)); + ASSERT_EQ(0, read_streamer->open(read_storage)); + size_t topk = 30; + ElapsedTime elapsed_time; + for (size_t i = 0; i < cnt; i += 1) { + NumericalVector vec(dim); + for (size_t j = 0; j < dim; ++j) { + vec[j] = i; + } + ctx->set_topk(topk); + ASSERT_EQ(0, read_streamer->search_impl(vec.data(), qmeta, ctx)); + // auto &result1 = ctx->result(); + // ASSERT_EQ(topk, result1.size()); + // ASSERT_EQ(i, result1[0].key()); + + // for (size_t j = 0; j < dim; ++j) { + // vec[j] = i + 0.1f; + // } + // ctx->set_topk(topk); + // ASSERT_EQ(0, read_streamer->search_impl(vec.data(), qmeta, ctx)); + // auto &result2 = ctx->result(); + // ASSERT_EQ(topk, result2.size()); + // ASSERT_EQ(i, result2[0].key()); + // ASSERT_EQ(i == cnt - 1 ? i - 1 : i + 1, result2[1].key()); + // ASSERT_EQ(i == 0 ? 2 : (i == cnt - 1 ? i - 2 : i - 1), result2[2].key()); + } + cout << "Elapsed time: " << elapsed_time.micro_seconds() << " us" << endl; + for (size_t i = 0; i < cnt; i += 1) { + NumericalVector vec(dim); + for (size_t j = 0; j < dim; ++j) { + vec[j] = i; + } + ctx->set_topk(topk); + ASSERT_EQ(0, read_streamer->search_impl(vec.data(), qmeta, ctx)); + // auto &result1 = ctx->result(); + // ASSERT_EQ(topk, result1.size()); + // ASSERT_EQ(i, result1[0].key()); + + // for (size_t j = 0; j < dim; ++j) { + // vec[j] = i + 0.1f; + // } + // ctx->set_topk(topk); + // ASSERT_EQ(0, read_streamer->search_impl(vec.data(), qmeta, ctx)); + // auto &result2 = ctx->result(); + // ASSERT_EQ(topk, result2.size()); + // ASSERT_EQ(i, result2[0].key()); + // ASSERT_EQ(i == cnt - 1 ? i - 1 : i + 1, result2[1].key()); + // ASSERT_EQ(i == 0 ? 2 : (i == cnt - 1 ? i - 2 : i - 1), result2[2].key()); + } + cout << "Elapsed time: " << elapsed_time.micro_seconds() << " us" << endl; read_streamer->close(); read_streamer.reset(); - // cout << "Elapsed time: " << elapsed_time.milli_seconds() << " ms" << endl; } #if defined(__GNUC__) || defined(__GNUG__) From 7df2716d2ac969e665422b2d4a85ae51cc3d47cf Mon Sep 17 00:00:00 2001 From: "yinzefeng.yzf" Date: Mon, 9 Feb 2026 15:33:34 +0800 Subject: [PATCH 04/11] upd buffer pool --- src/ailego/buffer/buffer_pool.cc | 75 +-- src/core/utility/buffer1_storage.cc | 438 ------------------ src/core/utility/buffer_storage.cc | 130 +++--- ..._test.cpp => hnsw_streamer_buffer_test.cc} | 0 4 files changed, 91 insertions(+), 552 deletions(-) delete mode 100644 src/core/utility/buffer1_storage.cc rename tests/core/algorithm/hnsw/{hnsw_streamer_buffer_test.cpp => hnsw_streamer_buffer_test.cc} (100%) diff --git a/src/ailego/buffer/buffer_pool.cc b/src/ailego/buffer/buffer_pool.cc index 061ead37..3ed461c1 100644 --- a/src/ailego/buffer/buffer_pool.cc +++ b/src/ailego/buffer/buffer_pool.cc @@ -1,23 +1,9 @@ #include +#include namespace zvec { namespace ailego { -void Counter::record(const std::string &name, int64_t value) { - auto it = static_counters.find(name); - if (it == static_counters.end()) { - auto counter = std::make_unique>(0); - it = static_counters.emplace(name, std::move(counter)).first; - } - it->second->fetch_add(value); -} - -void Counter::display() { - for (const auto &pair : static_counters) { - std::cout << pair.first << ": " << pair.second->load() << std::endl; - } -} - int LRUCache::init(size_t block_size) { block_size_ = block_size; for(size_t i = 0; i < CATCH_QUEUE_NUM; i++) { @@ -27,11 +13,9 @@ int LRUCache::init(size_t block_size) { } bool LRUCache::evict_single_block(BlockType &item) { - // std::cerr << "dequeue: " << item.first << std::endl; bool found = false; for(size_t i = 0; i < CATCH_QUEUE_NUM; i++) { found = queues_[i].try_dequeue(item); - // std::cerr << "dequeue: " << found << std::endl; if(found) { break; } @@ -41,7 +25,8 @@ bool LRUCache::evict_single_block(BlockType &item) { bool LRUCache::add_single_block(const LPMap *lp_map, const BlockType &block, int block_type) { bool ok = queues_[block_type].try_enqueue(block); - if(++evict_queue_insertions_ % block_size_ == 0) { + evict_queue_insertions_.fetch_add(1, std::memory_order_relaxed); + if(evict_queue_insertions_ % block_size_ == 0) { this->clear_dead_node(lp_map); } return ok; @@ -49,10 +34,14 @@ bool LRUCache::add_single_block(const LPMap *lp_map, const BlockType &block, int void LRUCache::clear_dead_node(const LPMap *lp_map) { for(int i = 0; i < CATCH_QUEUE_NUM; i++) { + int clear_size = block_size_ * 2; + if (queues_[i].size_approx() < clear_size * 4) { + continue; + } int clear_count = 0; ConcurrentQueue tmp(block_size_); BlockType item; - while(queues_[i].try_dequeue(item) && (clear_count++ < block_size_)) { + while(queues_[i].try_dequeue(item) && (clear_count++ < clear_size)) { if(!lp_map->isDeadBlock(item)) { tmp.try_enqueue(item); } @@ -82,14 +71,11 @@ void LPMap::init(size_t entry_num) { char* LPMap::acquire_block(block_id_t block_id) { assert(block_id < entry_num_); Entry &entry = entries_[block_id]; - if (entry.ref_count.load() == 0) { - ++entry.load_count; - // std::cout << entry.load_count.load() << std::endl; + if (entry.ref_count.load(std::memory_order_relaxed) == 0) { + entry.load_count.fetch_add(1, std::memory_order_relaxed); } - ++entry.ref_count; - // std::cout << entry.ref_count.load() << std::endl; - if (entry.ref_count.load() < 0) { - // std::cout << "acquire block failed: " << block_id << ", " << entry.ref_count.load() << std::endl; + entry.ref_count.fetch_add(1, std::memory_order_relaxed); + if (entry.ref_count.load(std::memory_order_relaxed) < 0) { return nullptr; } return entry.buffer; @@ -98,10 +84,9 @@ char* LPMap::acquire_block(block_id_t block_id) { void LPMap::release_block(block_id_t block_id) { assert(block_id < entry_num_); Entry &entry = entries_[block_id]; - int rc = entry.ref_count.fetch_sub(1); - // std::cout << "release block: " << block_id << ", " << entry.ref_count.load() << std::endl; - // assert(rc > 0); - if(entry.ref_count.load() == 0) { + + if (entry.ref_count.fetch_sub(1, std::memory_order_release) == 1) { + std::atomic_thread_fence(std::memory_order_acquire); LRUCache::BlockType block; block.first = block_id; block.second = entry.load_count.load(); @@ -110,7 +95,6 @@ void LPMap::release_block(block_id_t block_id) { } char* LPMap::evict_block(block_id_t block_id) { - // std::cout << "evict block: " << block_id << std::endl; assert(block_id < entry_num_); Entry &entry = entries_[block_id]; int expected = 0; @@ -127,15 +111,13 @@ char* LPMap::evict_block(block_id_t block_id) { char* LPMap::set_block_acquired(block_id_t block_id, char *buffer) { assert(block_id < entry_num_); Entry &entry = entries_[block_id]; - if (entry.ref_count.load() >= 0) { - entry.ref_count.fetch_add(1); - // std::cout << "Set block2 " << block_id << std::endl; + if (entry.ref_count.load(std::memory_order_relaxed) >= 0) { + entry.ref_count.fetch_add(1, std::memory_order_relaxed); return entry.buffer; } - // if (buffer == nullptr) std::cout << "Set block " << block_id << std::endl; entry.buffer = buffer; - entry.ref_count.store(1); - entry.load_count.fetch_add(1); + entry.ref_count.store(1, std::memory_order_relaxed); + entry.load_count.fetch_add(1, std::memory_order_relaxed); return buffer; } @@ -147,7 +129,6 @@ void LPMap::recycle(moodycamel::ConcurrentQueue &free_buffers) { return; } } while(isDeadBlock(block)); - // std::cout << "evict_block done: " << block.first << ", " << block.second << std::endl; char *buffer = evict_block(block.first); if (buffer) { free_buffers.try_enqueue(buffer); @@ -173,11 +154,9 @@ VecBufferPool::VecBufferPool(const std::string &filename, size_t pool_capacity, char *buffer = (char *)aligned_alloc(64, block_size); if (buffer != nullptr) { bool ok = free_buffers_.try_enqueue(buffer); - // if(!ok) std::cerr << i << std::endl; } } - std::cout << "buffer_num: " << buffer_num << std::endl; - std::cout << "entry_num: " << lp_map_.entry_num() << std::endl; + LOG_DEBUG("Buffer pool num: %zu, entry num: %zu", buffer_num, lp_map_.entry_num()); } VecBufferPoolHandle VecBufferPool::get_handle() { @@ -190,30 +169,26 @@ char* VecBufferPool::acquire_buffer(block_id_t block_id, size_t offset, size_t s return buffer; } { - // std::cerr << "block_id: " << block_id << ", offset: " << offset << ", size: " << size << std::endl; - // std::lock_guard lock(mutex_); bool found = free_buffers_.try_dequeue(buffer); - // std::cerr << "dequeue: " << found << std::endl; if (!found) { for (int i = 0; i < retry; i++) { lp_map_.recycle(free_buffers_); found = free_buffers_.try_dequeue(buffer); - // std::cerr << "dequeue: " << i << std::endl; if (found) { break; } } } if (!found) { - std::cerr << "Failed to get free buffer " << std::endl; + LOG_ERROR("Buffer pool failed to get free buffer"); return nullptr; } } ssize_t read_bytes = pread(fd_, buffer, size, offset); if (read_bytes != static_cast(size)) { - std::cerr << "Failed to read file at offset " << offset << std::endl; - exit(-1); + LOG_ERROR("Buffer pool failed to read file at offset: %zu", offset); + return nullptr; } char *placed_buffer = nullptr; { @@ -230,8 +205,8 @@ char* VecBufferPool::acquire_buffer(block_id_t block_id, size_t offset, size_t s int VecBufferPool::get_meta(size_t offset, size_t length, char *buffer) { ssize_t read_bytes = pread(fd_, buffer, length, offset); if (read_bytes != static_cast(length)) { - std::cerr << "Failed to read file at offset " << offset << std::endl; - exit(-1); + LOG_ERROR("Buffer pool failed to read file at offset: %zu", offset);LOG_ERROR("Buffer pool failed to read file at offset: %zu", offset); + return -1; } return 0; } diff --git a/src/core/utility/buffer1_storage.cc b/src/core/utility/buffer1_storage.cc deleted file mode 100644 index 1c582198..00000000 --- a/src/core/utility/buffer1_storage.cc +++ /dev/null @@ -1,438 +0,0 @@ -// Copyright 2025-present the zvec project -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -#include -// #include -#include -#include -#include -#include -#include -#include "utility_params.h" - -#include - -namespace zvec { -namespace core { - -/*! MMap File Storage - */ -class Buffer1Storage : public IndexStorage { - public: - /*! Index Storage Segment - */ - class Segment : public IndexStorage::Segment, - public std::enable_shared_from_this { - public: - //! Index Storage Pointer - typedef std::shared_ptr Pointer; - - //! Constructor - Segment(Buffer1Storage *owner, IndexMapping::Segment *segment, size_t segment_id) - : segment_(segment), - owner_(owner), - segment_id_(segment_id), - capacity_(static_cast(segment->meta()->data_size + - segment->meta()->padding_size)) {} - - //! Destructor - virtual ~Segment(void) {} - - //! Retrieve size of data - size_t data_size(void) const override { - return static_cast(segment_->meta()->data_size); - } - - //! Retrieve crc of data - uint32_t data_crc(void) const override { - return segment_->meta()->data_crc; - } - - //! Retrieve size of padding - size_t padding_size(void) const override { - return static_cast(segment_->meta()->padding_size); - } - - //! Retrieve capacity of segment - size_t capacity(void) const override { - return capacity_; - } - - //! Fetch data from segment (with own buffer) - size_t fetch(size_t offset, void *buf, size_t len) const override { - if (ailego_unlikely(offset + len > segment_->meta()->data_size)) { - auto meta = segment_->meta(); - if (offset > meta->data_size) { - offset = meta->data_size; - } - len = meta->data_size - offset; - } - memmove(buf, (const uint8_t *)(owner_->get_buffer(offset, len, segment_id_)) + offset, - len); - return len; - } - - //! Read data from segment - size_t read(size_t offset, const void **data, size_t len) override { - if (ailego_unlikely(offset + len > segment_->meta()->data_size)) { - auto meta = segment_->meta(); - if (offset > meta->data_size) { - offset = meta->data_size; - } - len = meta->data_size - offset; - } - size_t segment_offset = segment_->meta()->data_index + owner_->get_context_offset(); - *data = owner_->get_buffer(segment_offset, capacity_, segment_id_) + offset; - return len; - } - - size_t read(size_t offset, MemoryBlock &data, size_t len) override { - if (ailego_unlikely(offset + len > segment_->meta()->data_size)) { - auto meta = segment_->meta(); - if (offset > meta->data_size) { - offset = meta->data_size; - } - len = meta->data_size - offset; - } - size_t segment_offset = segment_->meta()->data_index + owner_->get_context_offset(); - data.reset(owner_->buffer_pool_handle_.get(), segment_id_, owner_->get_buffer(segment_offset, capacity_, segment_id_) + offset); - // data.reset(owner_->get_buffer(segment_offset, capacity_, segment_id_) + offset); - if (data.data()) { - return len; - } else { - LOG_ERROR("read error."); - return -1; - } - } - - //! Write data into the storage with offset - size_t write(size_t /*offset*/, const void * /*data*/, - size_t len) override { - return len; - } - - //! Resize size of data - size_t resize(size_t /*size*/) override { - return 0; - } - - //! Update crc of data - void update_data_crc(uint32_t /*crc*/) override {} - - //! Clone the segment - IndexStorage::Segment::Pointer clone(void) override { - return shared_from_this(); - } - - private: - IndexMapping::Segment *segment_{}; - Buffer1Storage *owner_{nullptr}; - size_t segment_id_{}; - size_t capacity_{}; - }; - - //! Destructor - virtual ~Buffer1Storage(void) { - this->cleanup(); - } - - //! Initialize storage - int init(const ailego::Params & /*params*/) override { - return 0; - } - - //! Cleanup storage - int cleanup(void) override { - this->close_index(); - return 0; - } - - //! Open storage - int open(const std::string &path, bool /*create*/) override { - LOG_INFO("open buffer storage 1"); - file_name_ = path; - buffer_pool_ = std::make_shared(path, 10u * 1024 * 1024 * 1024, 2490368 * 2); - buffer_pool_handle_ = - std::make_shared(buffer_pool_->get_handle()); - int ret = ParseToMapping(); - LOG_ERROR("segment count: %lu, max_segment_size: %lu", segments_.size(), max_segment_size_); - if(ret != 0) { - return ret; - } - return 0; - } - - char *get_buffer(size_t offset, size_t length, size_t block_id) { - return buffer_pool_handle_->get_block(offset, length, block_id); - } - - int get_meta(size_t offset, size_t length, char *out) { - return buffer_pool_handle_->get_meta(offset, length, out); - } - - int ParseHeader(size_t offset) { - char *buffer = new char[sizeof(header_)]; - get_meta(offset, sizeof(header_), buffer); - uint8_t *header_ptr = reinterpret_cast(buffer); - memcpy(&header_, header_ptr, sizeof(header_)); - delete[] buffer; - if (header_.meta_header_size != sizeof(IndexFormat::MetaHeader)) { - LOG_ERROR("Header meta size is invalid."); - return IndexError_InvalidLength; - } - if (ailego::Crc32c::Hash(&header_, sizeof(header_), header_.header_crc) != - header_.header_crc) { - LOG_ERROR("Header meta checksum is invalid."); - return IndexError_InvalidChecksum; - } - return 0; - } - - int ParseFooter(size_t offset) { - char *buffer = new char[sizeof(footer_)]; - get_meta(offset, sizeof(footer_), buffer); - uint8_t *footer_ptr = reinterpret_cast(buffer); - memcpy(&footer_, footer_ptr, sizeof(footer_)); - delete[] buffer; - if (offset < (size_t)footer_.segments_meta_size) { - LOG_ERROR("Footer meta size is invalid."); - return IndexError_InvalidLength; - } - if (ailego::Crc32c::Hash(&footer_, sizeof(footer_), footer_.footer_crc) != - footer_.footer_crc) { - LOG_ERROR("Footer meta checksum is invalid."); - return IndexError_InvalidChecksum; - } - return 0; - } - - int ParseSegment(size_t offset) { - segment_buffer_ = std::make_unique(footer_.segments_meta_size); - get_meta(offset, footer_.segments_meta_size, segment_buffer_.get()); - if (ailego::Crc32c::Hash(segment_buffer_.get(), footer_.segments_meta_size, 0u) != - footer_.segments_meta_crc) { - LOG_ERROR("Index segments meta checksum is invalid."); - return IndexError_InvalidChecksum; - } - IndexFormat::SegmentMeta *segment_start = - reinterpret_cast(segment_buffer_.get()); - uint32_t segment_ids_offset = footer_.segments_meta_size; - for (IndexFormat::SegmentMeta *iter = segment_start, - *end = segment_start + footer_.segment_count; - iter != end; ++iter) { - if (iter->segment_id_offset > footer_.segments_meta_size) { - return IndexError_InvalidValue; - } - if (iter->data_index > footer_.content_size) { - return IndexError_InvalidValue; - } - if (iter->data_index + iter->data_size > footer_.content_size) { - return IndexError_InvalidLength; - } - - if (iter->segment_id_offset < segment_ids_offset) { - segment_ids_offset = iter->segment_id_offset; - } - id_hash_.emplace( - std::string(reinterpret_cast(segment_start) + - iter->segment_id_offset), - segments_.size()); - segments_.emplace( - std::string(reinterpret_cast(segment_start) + - iter->segment_id_offset), - iter); - max_segment_size_ = std::max(max_segment_size_, iter->data_size + iter->padding_size); - if (sizeof(IndexFormat::SegmentMeta) * footer_.segment_count > - footer_.segments_meta_size) { - return IndexError_InvalidLength; - } - } - return 0; - } - - int ParseToMapping() { - ParseHeader(0); - // Unpack footer - if (header_.meta_footer_size != sizeof(IndexFormat::MetaFooter)) { - return IndexError_InvalidLength; - } - if ((int32_t)header_.meta_footer_offset < 0) { - return IndexError_Unsupported; - } - size_t footer_offset = header_.meta_footer_offset; - ParseFooter(footer_offset); - - // Unpack segment table - if (sizeof(IndexFormat::SegmentMeta) * footer_.segment_count > - footer_.segments_meta_size) { - return IndexError_InvalidLength; - } - const size_t segment_start_offset = footer_offset - footer_.segments_meta_size; - ParseSegment(segment_start_offset); - return 0; - } - - //! Flush storage - int flush(void) override { - return this->flush_index(); - } - - //! Close storage - int close(void) override { - this->close_index(); - return 0; - } - - //! Append a segment into storage - int append(const std::string &id, size_t size) override { - return this->append_segment(id, size); - } - - //! Refresh meta information (checksum, update time, etc.) - void refresh(uint64_t chkp) override { - this->refresh_index(chkp); - } - - //! Retrieve check point of storage - uint64_t check_point(void) const override { - return footer_.check_point; - } - - //! Retrieve a segment by id - IndexStorage::Segment::Pointer get(const std::string &id, int) override { - IndexMapping::Segment *segment = this->get_segment(id); - if (!segment) { - return Buffer1Storage::Segment::Pointer(); - } - return std::make_shared(this, segment, - id_hash_[id]); - } - - //! Test if it a segment exists - bool has(const std::string &id) const override { - return this->has_segment(id); - } - - //! Retrieve magic number of index - uint32_t magic(void) const override { - return header_.magic; - } - - uint32_t get_context_offset() { - return header_.content_offset; - } - - protected: - //! Initialize index version segment - int init_version_segment(void) { - size_t data_size = std::strlen(IndexVersion::Details()); - int error_code = - this->append_segment(INDEX_VERSION_SEGMENT_NAME, data_size); - if (error_code != 0) { - return error_code; - } - - IndexMapping::Segment *segment = get_segment(INDEX_VERSION_SEGMENT_NAME); - if (!segment) { - return IndexError_MMapFile; - } - auto meta = segment->meta(); - size_t capacity = static_cast(meta->padding_size + meta->data_size); - memcpy(segment->data(), IndexVersion::Details(), data_size); - segment->set_dirty(); - meta->data_crc = ailego::Crc32c::Hash(segment->data(), data_size, 0); - meta->data_size = data_size; - meta->padding_size = capacity - data_size; - return 0; - } - - //! Initialize index file - int init_index(const std::string &path) { - // Add index version - int error_code = this->init_version_segment(); - if (error_code != 0) { - return error_code; - } - - // Refresh mapping - this->refresh_index(0); - return 0; - } - - //! Set the index file as dirty - void set_as_dirty(void) { - index_dirty_ = true; - } - - //! Refresh meta information (checksum, update time, etc.) - void refresh_index(uint64_t /*chkp*/) {} - - //! Flush index storage - int flush_index(void) { - return 0; - } - - //! Close index storage - void close_index(void) { - std::lock_guard latch(mapping_mutex_); - file_name_.clear(); - segments_.clear(); - memset(&header_, 0, sizeof(header_)); - memset(&footer_, 0, sizeof(footer_)); - segment_buffer_.release(); - } - - //! Append a segment into storage - int append_segment(const std::string & /*id*/, size_t /*size*/) { - return 0; - } - - //! Test if a segment exists - bool has_segment(const std::string &id) const { - std::lock_guard latch(mapping_mutex_); - return (segments_.find(id) != segments_.end()); - } - - //! Get a segment from storage - IndexMapping::Segment *get_segment(const std::string &id) { - std::lock_guard latch(mapping_mutex_); - auto iter = segments_.find(id); - if (iter == segments_.end()) { - return nullptr; - } - IndexMapping::Segment *item = &iter->second; - return item; - } - - private: - bool index_dirty_{false}; - mutable std::mutex mapping_mutex_{}; - - // buffer manager - std::string file_name_; - IndexFormat::MetaHeader header_; - IndexFormat::MetaFooter footer_; - std::map segments_{}; - std::map id_hash_{}; - size_t max_segment_size_{0}; - std::unique_ptr segment_buffer_{nullptr}; - - ailego::VecBufferPool::Pointer buffer_pool_{nullptr}; - ailego::VecBufferPoolHandle::Pointer buffer_pool_handle_{nullptr}; -}; - -INDEX_FACTORY_REGISTER_STORAGE_ALIAS(BufferStorage, Buffer1Storage); - -} // namespace core -} // namespace zvec \ No newline at end of file diff --git a/src/core/utility/buffer_storage.cc b/src/core/utility/buffer_storage.cc index d4b23c87..13aee16a 100644 --- a/src/core/utility/buffer_storage.cc +++ b/src/core/utility/buffer_storage.cc @@ -13,13 +13,16 @@ // limitations under the License. #include -#include +// #include +#include #include #include #include #include #include "utility_params.h" +#include + namespace zvec { namespace core { @@ -36,9 +39,10 @@ class BufferStorage : public IndexStorage { typedef std::shared_ptr Pointer; //! Constructor - Segment(BufferStorage *owner, IndexMapping::Segment *segment) + Segment(BufferStorage *owner, IndexMapping::Segment *segment, size_t segment_id) : segment_(segment), owner_(owner), + segment_id_(segment_id), capacity_(static_cast(segment->meta()->data_size + segment->meta()->padding_size)) {} @@ -74,9 +78,7 @@ class BufferStorage : public IndexStorage { } len = meta->data_size - offset; } - ailego::BufferHandle buffer_handle = - owner_->get_buffer_handle(offset, len); - memmove(buf, (const uint8_t *)buffer_handle.pin_vector_data() + offset, + memmove(buf, (const uint8_t *)(owner_->get_buffer(offset, len, segment_id_)) + offset, len); return len; } @@ -90,11 +92,8 @@ class BufferStorage : public IndexStorage { } len = meta->data_size - offset; } - size_t buffer_offset = - segment_->meta()->data_index + owner_->get_context_offset() + offset; - ailego::BufferHandle buffer_handle = - owner_->get_buffer_handle(buffer_offset, len); - *data = buffer_handle.pin_vector_data(); + size_t segment_offset = segment_->meta()->data_index + owner_->get_context_offset(); + *data = owner_->get_buffer(segment_offset, capacity_, segment_id_) + offset; return len; } @@ -106,16 +105,13 @@ class BufferStorage : public IndexStorage { } len = meta->data_size - offset; } - size_t buffer_offset = - segment_->meta()->data_index + owner_->get_context_offset() + offset; - data.reset(owner_->get_buffer_handle_ptr(buffer_offset, len)); + size_t segment_offset = segment_->meta()->data_index + owner_->get_context_offset(); + data.reset(owner_->buffer_pool_handle_.get(), segment_id_, owner_->get_buffer(segment_offset, capacity_, segment_id_) + offset); + // data.reset(owner_->get_buffer(segment_offset, capacity_, segment_id_) + offset); if (data.data()) { return len; } else { - LOG_ERROR( - "Buffer handle is null, now used memory: %zu, new: %zu", - (size_t)ailego::BufferManager::Instance().total_size_in_bytes(), - len); + LOG_ERROR("read error."); return -1; } } @@ -142,6 +138,7 @@ class BufferStorage : public IndexStorage { private: IndexMapping::Segment *segment_{}; BufferStorage *owner_{nullptr}; + size_t segment_id_{}; size_t capacity_{}; }; @@ -163,29 +160,39 @@ class BufferStorage : public IndexStorage { //! Open storage int open(const std::string &path, bool /*create*/) override { + LOG_INFO("open buffer storage 1"); file_name_ = path; - return ParseToMapping(); + buffer_pool_ = std::make_shared(path, 20lu * 1024 * 1024 * 1024, 2490368 * 2); + buffer_pool_handle_ = + std::make_shared(buffer_pool_->get_handle()); + int ret = ParseToMapping(); + LOG_ERROR("segment count: %lu, max_segment_size: %lu", segments_.size(), max_segment_size_); + for(auto iter = segments_.begin(); iter != segments_.end(); iter++) { + auto seg = this->get(iter->first, 0); + MemoryBlock block; + int len = seg->read(0, block, 1); + LOG_ERROR("segment %s: %d", iter->first.c_str(), len); + } + if(ret != 0) { + return ret; + } + return 0; } - ailego::BufferHandle get_buffer_handle(int offset, int length) { - ailego::BufferID buffer_id = - ailego::BufferID::VectorID(file_name_, offset, length); - return ailego::BufferManager::Instance().acquire(buffer_id); + char *get_buffer(size_t offset, size_t length, size_t block_id) { + return buffer_pool_handle_->get_block(offset, length, block_id); } - ailego::BufferHandle::Pointer get_buffer_handle_ptr(int offset, int length) { - ailego::BufferID buffer_id = - ailego::BufferID::VectorID(file_name_, offset, length); - return ailego::BufferManager::Instance().acquire_ptr(buffer_id); + int get_meta(size_t offset, size_t length, char *out) { + return buffer_pool_handle_->get_meta(offset, length, out); } - int ParseHeader(int offset) { - ailego::BufferHandle header_handle = - get_buffer_handle(offset, sizeof(header_)); - void *buffer = header_handle.pin_vector_data(); + int ParseHeader(size_t offset) { + char *buffer = new char[sizeof(header_)]; + get_meta(offset, sizeof(header_), buffer); uint8_t *header_ptr = reinterpret_cast(buffer); memcpy(&header_, header_ptr, sizeof(header_)); - header_handle.unpin_vector_data(); + delete[] buffer; if (header_.meta_header_size != sizeof(IndexFormat::MetaHeader)) { LOG_ERROR("Header meta size is invalid."); return IndexError_InvalidLength; @@ -198,14 +205,13 @@ class BufferStorage : public IndexStorage { return 0; } - int ParseFooter(int offset) { - ailego::BufferHandle footer_handle = - get_buffer_handle(offset, sizeof(footer_)); - void *buffer = footer_handle.pin_vector_data(); + int ParseFooter(size_t offset) { + char *buffer = new char[sizeof(footer_)]; + get_meta(offset, sizeof(footer_), buffer); uint8_t *footer_ptr = reinterpret_cast(buffer); memcpy(&footer_, footer_ptr, sizeof(footer_)); - footer_handle.unpin_vector_data(); - if (offset < (int)footer_.segments_meta_size) { + delete[] buffer; + if (offset < (size_t)footer_.segments_meta_size) { LOG_ERROR("Footer meta size is invalid."); return IndexError_InvalidLength; } @@ -217,17 +223,16 @@ class BufferStorage : public IndexStorage { return 0; } - int ParseSegment(int offset) { - ailego::BufferHandle segment_start_handle = - get_buffer_handle(offset, footer_.segments_meta_size); - void *segment_buffer = segment_start_handle.pin_vector_data(); - if (ailego::Crc32c::Hash(segment_buffer, footer_.segments_meta_size, 0u) != + int ParseSegment(size_t offset) { + segment_buffer_ = std::make_unique(footer_.segments_meta_size); + get_meta(offset, footer_.segments_meta_size, segment_buffer_.get()); + if (ailego::Crc32c::Hash(segment_buffer_.get(), footer_.segments_meta_size, 0u) != footer_.segments_meta_crc) { LOG_ERROR("Index segments meta checksum is invalid."); return IndexError_InvalidChecksum; } IndexFormat::SegmentMeta *segment_start = - reinterpret_cast(segment_buffer); + reinterpret_cast(segment_buffer_.get()); uint32_t segment_ids_offset = footer_.segments_meta_size; for (IndexFormat::SegmentMeta *iter = segment_start, *end = segment_start + footer_.segment_count; @@ -245,10 +250,15 @@ class BufferStorage : public IndexStorage { if (iter->segment_id_offset < segment_ids_offset) { segment_ids_offset = iter->segment_id_offset; } + id_hash_.emplace( + std::string(reinterpret_cast(segment_start) + + iter->segment_id_offset), + segments_.size()); segments_.emplace( std::string(reinterpret_cast(segment_start) + iter->segment_id_offset), iter); + max_segment_size_ = std::max(max_segment_size_, iter->data_size + iter->padding_size); if (sizeof(IndexFormat::SegmentMeta) * footer_.segment_count > footer_.segments_meta_size) { return IndexError_InvalidLength; @@ -259,7 +269,6 @@ class BufferStorage : public IndexStorage { int ParseToMapping() { ParseHeader(0); - // Unpack footer if (header_.meta_footer_size != sizeof(IndexFormat::MetaFooter)) { return IndexError_InvalidLength; @@ -275,7 +284,7 @@ class BufferStorage : public IndexStorage { footer_.segments_meta_size) { return IndexError_InvalidLength; } - const int segment_start_offset = footer_offset - footer_.segments_meta_size; + const size_t segment_start_offset = footer_offset - footer_.segments_meta_size; ParseSegment(segment_start_offset); return 0; } @@ -312,7 +321,8 @@ class BufferStorage : public IndexStorage { if (!segment) { return BufferStorage::Segment::Pointer(); } - return std::make_shared(this, segment); + return std::make_shared(this, segment, + id_hash_[id]); } //! Test if it a segment exists @@ -355,22 +365,14 @@ class BufferStorage : public IndexStorage { //! Initialize index file int init_index(const std::string &path) { - int error_code = mapping_.create(path, segment_meta_capacity_); - if (error_code != 0) { - return error_code; - } - // Add index version - error_code = this->init_version_segment(); + int error_code = this->init_version_segment(); if (error_code != 0) { return error_code; } // Refresh mapping this->refresh_index(0); - - // Close mapping - mapping_.close(); return 0; } @@ -394,6 +396,7 @@ class BufferStorage : public IndexStorage { segments_.clear(); memset(&header_, 0, sizeof(header_)); memset(&footer_, 0, sizeof(footer_)); + segment_buffer_.release(); } //! Append a segment into storage @@ -419,14 +422,7 @@ class BufferStorage : public IndexStorage { } private: - // mmap - uint32_t segment_meta_capacity_{1024 * 1024}; - // bool copy_on_write_{false}; - // bool force_flush_{false}; - // bool memory_locked_{false}; - // bool memory_warmup_{false}; bool index_dirty_{false}; - mutable IndexMapping mapping_{}; mutable std::mutex mapping_mutex_{}; // buffer manager @@ -434,9 +430,15 @@ class BufferStorage : public IndexStorage { IndexFormat::MetaHeader header_; IndexFormat::MetaFooter footer_; std::map segments_{}; + std::map id_hash_{}; + size_t max_segment_size_{0}; + std::unique_ptr segment_buffer_{nullptr}; + + ailego::VecBufferPool::Pointer buffer_pool_{nullptr}; + ailego::VecBufferPoolHandle::Pointer buffer_pool_handle_{nullptr}; }; -// INDEX_FACTORY_REGISTER_STORAGE(BufferStorage); +INDEX_FACTORY_REGISTER_STORAGE(BufferStorage); } // namespace core -} // namespace zvec +} // namespace zvec \ No newline at end of file diff --git a/tests/core/algorithm/hnsw/hnsw_streamer_buffer_test.cpp b/tests/core/algorithm/hnsw/hnsw_streamer_buffer_test.cc similarity index 100% rename from tests/core/algorithm/hnsw/hnsw_streamer_buffer_test.cpp rename to tests/core/algorithm/hnsw/hnsw_streamer_buffer_test.cc From 11a0e475d154a57e54f23ae7791352db08e48d34 Mon Sep 17 00:00:00 2001 From: "yinzefeng.yzf" Date: Mon, 9 Feb 2026 19:10:42 +0800 Subject: [PATCH 05/11] clang format --- src/ailego/buffer/buffer_pool.cc | 384 +- src/core/utility/buffer_storage.cc | 38 +- src/include/zvec/ailego/buffer/buffer_pool.h | 40 +- .../zvec/ailego/buffer/concurrentqueue.h | 7693 +++++++++-------- 4 files changed, 4418 insertions(+), 3737 deletions(-) diff --git a/src/ailego/buffer/buffer_pool.cc b/src/ailego/buffer/buffer_pool.cc index 3ed461c1..81ed92bf 100644 --- a/src/ailego/buffer/buffer_pool.cc +++ b/src/ailego/buffer/buffer_pool.cc @@ -5,227 +5,233 @@ namespace zvec { namespace ailego { int LRUCache::init(size_t block_size) { - block_size_ = block_size; - for(size_t i = 0; i < CATCH_QUEUE_NUM; i++) { - queues_.push_back(ConcurrentQueue(block_size)); - } - return 0; + block_size_ = block_size; + for (size_t i = 0; i < CATCH_QUEUE_NUM; i++) { + queues_.push_back(ConcurrentQueue(block_size)); + } + return 0; } bool LRUCache::evict_single_block(BlockType &item) { - bool found = false; - for(size_t i = 0; i < CATCH_QUEUE_NUM; i++) { - found = queues_[i].try_dequeue(item); - if(found) { - break; - } - } - return found; -} - -bool LRUCache::add_single_block(const LPMap *lp_map, const BlockType &block, int block_type) { - bool ok = queues_[block_type].try_enqueue(block); - evict_queue_insertions_.fetch_add(1, std::memory_order_relaxed); - if(evict_queue_insertions_ % block_size_ == 0) { - this->clear_dead_node(lp_map); - } - return ok; + bool found = false; + for (size_t i = 0; i < CATCH_QUEUE_NUM; i++) { + found = queues_[i].try_dequeue(item); + if (found) { + break; + } + } + return found; +} + +bool LRUCache::add_single_block(const LPMap *lp_map, const BlockType &block, + int block_type) { + bool ok = queues_[block_type].try_enqueue(block); + evict_queue_insertions_.fetch_add(1, std::memory_order_relaxed); + if (evict_queue_insertions_ % block_size_ == 0) { + this->clear_dead_node(lp_map); + } + return ok; } void LRUCache::clear_dead_node(const LPMap *lp_map) { - for(int i = 0; i < CATCH_QUEUE_NUM; i++) { - int clear_size = block_size_ * 2; - if (queues_[i].size_approx() < clear_size * 4) { - continue; - } - int clear_count = 0; - ConcurrentQueue tmp(block_size_); - BlockType item; - while(queues_[i].try_dequeue(item) && (clear_count++ < clear_size)) { - if(!lp_map->isDeadBlock(item)) { - tmp.try_enqueue(item); - } - } - while(tmp.try_dequeue(item)) { - if(!lp_map->isDeadBlock(item)) { - queues_[i].try_enqueue(item); - } - } - } + for (int i = 0; i < CATCH_QUEUE_NUM; i++) { + int clear_size = block_size_ * 2; + if (queues_[i].size_approx() < clear_size * 4) { + continue; + } + int clear_count = 0; + ConcurrentQueue tmp(block_size_); + BlockType item; + while (queues_[i].try_dequeue(item) && (clear_count++ < clear_size)) { + if (!lp_map->isDeadBlock(item)) { + tmp.try_enqueue(item); + } + } + while (tmp.try_dequeue(item)) { + if (!lp_map->isDeadBlock(item)) { + queues_[i].try_enqueue(item); + } + } + } } void LPMap::init(size_t entry_num) { - if (entries_) { - delete[] entries_; - } - entry_num_ = entry_num; - entries_ = new Entry[entry_num_]; - for (size_t i = 0; i < entry_num_; i++) { - entries_[i].ref_count.store(std::numeric_limits::min()); - entries_[i].load_count.store(0); - entries_[i].buffer = nullptr; - } - cache_.init(entry_num); -} - -char* LPMap::acquire_block(block_id_t block_id) { - assert(block_id < entry_num_); - Entry &entry = entries_[block_id]; - if (entry.ref_count.load(std::memory_order_relaxed) == 0) { - entry.load_count.fetch_add(1, std::memory_order_relaxed); - } - entry.ref_count.fetch_add(1, std::memory_order_relaxed); - if (entry.ref_count.load(std::memory_order_relaxed) < 0) { - return nullptr; - } - return entry.buffer; + if (entries_) { + delete[] entries_; + } + entry_num_ = entry_num; + entries_ = new Entry[entry_num_]; + for (size_t i = 0; i < entry_num_; i++) { + entries_[i].ref_count.store(std::numeric_limits::min()); + entries_[i].load_count.store(0); + entries_[i].buffer = nullptr; + } + cache_.init(entry_num); +} + +char *LPMap::acquire_block(block_id_t block_id) { + assert(block_id < entry_num_); + Entry &entry = entries_[block_id]; + if (entry.ref_count.load(std::memory_order_relaxed) == 0) { + entry.load_count.fetch_add(1, std::memory_order_relaxed); + } + entry.ref_count.fetch_add(1, std::memory_order_relaxed); + if (entry.ref_count.load(std::memory_order_relaxed) < 0) { + return nullptr; + } + return entry.buffer; } void LPMap::release_block(block_id_t block_id) { - assert(block_id < entry_num_); - Entry &entry = entries_[block_id]; - - if (entry.ref_count.fetch_sub(1, std::memory_order_release) == 1) { - std::atomic_thread_fence(std::memory_order_acquire); - LRUCache::BlockType block; - block.first = block_id; - block.second = entry.load_count.load(); - cache_.add_single_block(this, block, 0); - } -} - -char* LPMap::evict_block(block_id_t block_id) { - assert(block_id < entry_num_); - Entry &entry = entries_[block_id]; - int expected = 0; - if (entry.ref_count.compare_exchange_strong( - expected, std::numeric_limits::min())) { - char *buffer = entry.buffer; - entry.buffer = nullptr; - return buffer; - } else { - return nullptr; - } -} - -char* LPMap::set_block_acquired(block_id_t block_id, char *buffer) { - assert(block_id < entry_num_); - Entry &entry = entries_[block_id]; - if (entry.ref_count.load(std::memory_order_relaxed) >= 0) { - entry.ref_count.fetch_add(1, std::memory_order_relaxed); - return entry.buffer; - } - entry.buffer = buffer; - entry.ref_count.store(1, std::memory_order_relaxed); - entry.load_count.fetch_add(1, std::memory_order_relaxed); - return buffer; + assert(block_id < entry_num_); + Entry &entry = entries_[block_id]; + + if (entry.ref_count.fetch_sub(1, std::memory_order_release) == 1) { + std::atomic_thread_fence(std::memory_order_acquire); + LRUCache::BlockType block; + block.first = block_id; + block.second = entry.load_count.load(); + cache_.add_single_block(this, block, 0); + } +} + +char *LPMap::evict_block(block_id_t block_id) { + assert(block_id < entry_num_); + Entry &entry = entries_[block_id]; + int expected = 0; + if (entry.ref_count.compare_exchange_strong( + expected, std::numeric_limits::min())) { + char *buffer = entry.buffer; + entry.buffer = nullptr; + return buffer; + } else { + return nullptr; + } +} + +char *LPMap::set_block_acquired(block_id_t block_id, char *buffer) { + assert(block_id < entry_num_); + Entry &entry = entries_[block_id]; + if (entry.ref_count.load(std::memory_order_relaxed) >= 0) { + entry.ref_count.fetch_add(1, std::memory_order_relaxed); + return entry.buffer; + } + entry.buffer = buffer; + entry.ref_count.store(1, std::memory_order_relaxed); + entry.load_count.fetch_add(1, std::memory_order_relaxed); + return buffer; } void LPMap::recycle(moodycamel::ConcurrentQueue &free_buffers) { - LRUCache::BlockType block; - do { - bool ok = cache_.evict_single_block(block); - if(!ok) { - return; - } - } while(isDeadBlock(block)); - char *buffer = evict_block(block.first); - if (buffer) { - free_buffers.try_enqueue(buffer); - } -} - -VecBufferPool::VecBufferPool(const std::string &filename, size_t pool_capacity, size_t block_size) - : pool_capacity_(pool_capacity) { - fd_ = open(filename.c_str(), O_RDONLY); - if (fd_ < 0) { - throw std::runtime_error("Failed to open file: " + filename); - } - struct stat st; - if (fstat(fd_, &st) < 0) { - throw std::runtime_error("Failed to stat file: " + filename); - } - file_size_ = st.st_size; - - size_t buffer_num = pool_capacity_ / block_size; - size_t block_num = file_size_ / block_size + 500; - lp_map_.init(block_num); - for (size_t i = 0; i < buffer_num; i++) { - char *buffer = (char *)aligned_alloc(64, block_size); - if (buffer != nullptr) { - bool ok = free_buffers_.try_enqueue(buffer); - } - } - LOG_DEBUG("Buffer pool num: %zu, entry num: %zu", buffer_num, lp_map_.entry_num()); + LRUCache::BlockType block; + do { + bool ok = cache_.evict_single_block(block); + if (!ok) { + return; + } + } while (isDeadBlock(block)); + char *buffer = evict_block(block.first); + if (buffer) { + free_buffers.try_enqueue(buffer); + } +} + +VecBufferPool::VecBufferPool(const std::string &filename, size_t pool_capacity, + size_t block_size) + : pool_capacity_(pool_capacity) { + fd_ = open(filename.c_str(), O_RDONLY); + if (fd_ < 0) { + throw std::runtime_error("Failed to open file: " + filename); + } + struct stat st; + if (fstat(fd_, &st) < 0) { + throw std::runtime_error("Failed to stat file: " + filename); + } + file_size_ = st.st_size; + + size_t buffer_num = pool_capacity_ / block_size; + size_t block_num = file_size_ / block_size + 500; + lp_map_.init(block_num); + for (size_t i = 0; i < buffer_num; i++) { + char *buffer = (char *)aligned_alloc(64, block_size); + if (buffer != nullptr) { + bool ok = free_buffers_.try_enqueue(buffer); + } + } + LOG_DEBUG("Buffer pool num: %zu, entry num: %zu", buffer_num, + lp_map_.entry_num()); } VecBufferPoolHandle VecBufferPool::get_handle() { - return VecBufferPoolHandle(*this); -} - -char* VecBufferPool::acquire_buffer(block_id_t block_id, size_t offset, size_t size, int retry) { - char *buffer = lp_map_.acquire_block(block_id); - if (buffer) { - return buffer; - } - { - bool found = free_buffers_.try_dequeue(buffer); - if (!found) { - for (int i = 0; i < retry; i++) { - lp_map_.recycle(free_buffers_); - found = free_buffers_.try_dequeue(buffer); - if (found) { - break; - } - } - } - if (!found) { - LOG_ERROR("Buffer pool failed to get free buffer"); - return nullptr; - } - } - - ssize_t read_bytes = pread(fd_, buffer, size, offset); - if (read_bytes != static_cast(size)) { - LOG_ERROR("Buffer pool failed to read file at offset: %zu", offset); - return nullptr; - } - char *placed_buffer = nullptr; - { - std::lock_guard lock(mutex_); - placed_buffer = lp_map_.set_block_acquired(block_id, buffer); - } - if (placed_buffer != buffer) { - // another thread has set the block - free_buffers_.try_enqueue(buffer); - } - return placed_buffer; + return VecBufferPoolHandle(*this); +} + +char *VecBufferPool::acquire_buffer(block_id_t block_id, size_t offset, + size_t size, int retry) { + char *buffer = lp_map_.acquire_block(block_id); + if (buffer) { + return buffer; + } + { + bool found = free_buffers_.try_dequeue(buffer); + if (!found) { + for (int i = 0; i < retry; i++) { + lp_map_.recycle(free_buffers_); + found = free_buffers_.try_dequeue(buffer); + if (found) { + break; + } + } + } + if (!found) { + LOG_ERROR("Buffer pool failed to get free buffer"); + return nullptr; + } + } + + ssize_t read_bytes = pread(fd_, buffer, size, offset); + if (read_bytes != static_cast(size)) { + LOG_ERROR("Buffer pool failed to read file at offset: %zu", offset); + return nullptr; + } + char *placed_buffer = nullptr; + { + std::lock_guard lock(mutex_); + placed_buffer = lp_map_.set_block_acquired(block_id, buffer); + } + if (placed_buffer != buffer) { + // another thread has set the block + free_buffers_.try_enqueue(buffer); + } + return placed_buffer; } int VecBufferPool::get_meta(size_t offset, size_t length, char *buffer) { - ssize_t read_bytes = pread(fd_, buffer, length, offset); - if (read_bytes != static_cast(length)) { - LOG_ERROR("Buffer pool failed to read file at offset: %zu", offset);LOG_ERROR("Buffer pool failed to read file at offset: %zu", offset); - return -1; - } - return 0; + ssize_t read_bytes = pread(fd_, buffer, length, offset); + if (read_bytes != static_cast(length)) { + LOG_ERROR("Buffer pool failed to read file at offset: %zu", offset); + LOG_ERROR("Buffer pool failed to read file at offset: %zu", offset); + return -1; + } + return 0; } -char* VecBufferPoolHandle::get_block(size_t offset, size_t size, size_t block_id) { - char *buffer = pool.acquire_buffer(block_id, offset, size, 5); - return buffer; +char *VecBufferPoolHandle::get_block(size_t offset, size_t size, + size_t block_id) { + char *buffer = pool.acquire_buffer(block_id, offset, size, 5); + return buffer; } int VecBufferPoolHandle::get_meta(size_t offset, size_t length, char *buffer) { - return pool.get_meta(offset, length, buffer); + return pool.get_meta(offset, length, buffer); } void VecBufferPoolHandle::release_one(block_id_t block_id) { - pool.lp_map_.release_block(block_id); + pool.lp_map_.release_block(block_id); } void VecBufferPoolHandle::acquire_one(block_id_t block_id) { - pool.lp_map_.acquire_block(block_id); + pool.lp_map_.acquire_block(block_id); } } // namespace ailego diff --git a/src/core/utility/buffer_storage.cc b/src/core/utility/buffer_storage.cc index 3765fd15..dcdb13d3 100644 --- a/src/core/utility/buffer_storage.cc +++ b/src/core/utility/buffer_storage.cc @@ -15,14 +15,13 @@ #include // #include #include +#include #include #include #include #include #include "utility_params.h" -#include - namespace zvec { namespace core { @@ -81,7 +80,9 @@ class BufferStorage : public IndexStorage { } len = meta->data_size - offset; } - memmove(buf, (const uint8_t *)(owner_->get_buffer(offset, len, segment_id_)) + offset, + memmove(buf, + (const uint8_t *)(owner_->get_buffer(offset, len, segment_id_)) + + offset, len); return len; } @@ -98,7 +99,8 @@ class BufferStorage : public IndexStorage { size_t buffer_offset = segment_header_start_offset_ + segment_header_->content_offset + segment_->meta()->data_index + offset; - *data = owner_->get_buffer(buffer_offset, capacity_, segment_id_) + offset; + *data = + owner_->get_buffer(buffer_offset, capacity_, segment_id_) + offset; return len; } @@ -113,8 +115,11 @@ class BufferStorage : public IndexStorage { size_t buffer_offset = segment_header_start_offset_ + segment_header_->content_offset + segment_->meta()->data_index + offset; - data.reset(owner_->buffer_pool_handle_.get(), segment_id_, owner_->get_buffer(buffer_offset, capacity_, segment_id_) + offset); - // data.reset(owner_->get_buffer(buffer_offset, capacity_, segment_id_) + offset); + data.reset( + owner_->buffer_pool_handle_.get(), segment_id_, + owner_->get_buffer(buffer_offset, capacity_, segment_id_) + offset); + // data.reset(owner_->get_buffer(buffer_offset, capacity_, segment_id_) + + // offset); if (data.data()) { return len; } else { @@ -174,18 +179,20 @@ class BufferStorage : public IndexStorage { int open(const std::string &path, bool /*create*/) override { LOG_INFO("open buffer storage 1"); file_name_ = path; - buffer_pool_ = std::make_shared(path, 20lu * 1024 * 1024 * 1024, 2490368 * 2); - buffer_pool_handle_ = - std::make_shared(buffer_pool_->get_handle()); + buffer_pool_ = std::make_shared( + path, 20lu * 1024 * 1024 * 1024, 2490368 * 2); + buffer_pool_handle_ = std::make_shared( + buffer_pool_->get_handle()); int ret = ParseToMapping(); - LOG_ERROR("segment count: %lu, max_segment_size: %lu", segments_.size(), max_segment_size_); - for(auto iter = segments_.begin(); iter != segments_.end(); iter++) { + LOG_ERROR("segment count: %lu, max_segment_size: %lu", segments_.size(), + max_segment_size_); + for (auto iter = segments_.begin(); iter != segments_.end(); iter++) { auto seg = this->get(iter->first, 0); MemoryBlock block; int len = seg->read(0, block, 1); LOG_ERROR("segment %s: %d", iter->first.c_str(), len); } - if(ret != 0) { + if (ret != 0) { return ret; } return 0; @@ -238,8 +245,8 @@ class BufferStorage : public IndexStorage { int ParseSegment(size_t offset) { segment_buffer_ = std::make_unique(footer_.segments_meta_size); get_meta(offset, footer_.segments_meta_size, segment_buffer_.get()); - if (ailego::Crc32c::Hash(segment_buffer_.get(), footer_.segments_meta_size, 0u) != - footer_.segments_meta_crc) { + if (ailego::Crc32c::Hash(segment_buffer_.get(), footer_.segments_meta_size, + 0u) != footer_.segments_meta_crc) { LOG_ERROR("Index segments meta checksum is invalid."); return IndexError_InvalidChecksum; } @@ -271,7 +278,8 @@ class BufferStorage : public IndexStorage { iter->segment_id_offset), IndexMapping::SegmentInfo{IndexMapping::Segment{iter}, current_header_start_offset_, &header_}); - max_segment_size_ = std::max(max_segment_size_, iter->data_size + iter->padding_size); + max_segment_size_ = + std::max(max_segment_size_, iter->data_size + iter->padding_size); if (sizeof(IndexFormat::SegmentMeta) * footer_.segment_count > footer_.segments_meta_size) { return IndexError_InvalidLength; diff --git a/src/include/zvec/ailego/buffer/buffer_pool.h b/src/include/zvec/ailego/buffer/buffer_pool.h index 34c69d51..f1a0149c 100644 --- a/src/include/zvec/ailego/buffer/buffer_pool.h +++ b/src/include/zvec/ailego/buffer/buffer_pool.h @@ -11,12 +11,12 @@ #include #include #include +#include #include #include #include #include #include -#include #include "concurrentqueue.h" namespace zvec { @@ -28,23 +28,24 @@ using version_t = size_t; class LPMap; class LRUCache { - public: - typedef std::pair BlockType; - typedef moodycamel::ConcurrentQueue ConcurrentQueue; + public: + typedef std::pair BlockType; + typedef moodycamel::ConcurrentQueue ConcurrentQueue; - int init(size_t block_size); + int init(size_t block_size); - bool evict_single_block(BlockType &item); + bool evict_single_block(BlockType &item); - bool add_single_block(const LPMap *lp_map, const BlockType &block, int block_type); + bool add_single_block(const LPMap *lp_map, const BlockType &block, + int block_type); - void clear_dead_node(const LPMap *lp_map); + void clear_dead_node(const LPMap *lp_map); - private: - constexpr static size_t CATCH_QUEUE_NUM = 3; - int block_size_; - std::vector queues_; - alignas(64) std::atomic evict_queue_insertions_{0}; + private: + constexpr static size_t CATCH_QUEUE_NUM = 3; + int block_size_; + std::vector queues_; + alignas(64) std::atomic evict_queue_insertions_{0}; }; class LPMap { @@ -95,15 +96,17 @@ class VecBufferPoolHandle; class VecBufferPool { public: typedef std::shared_ptr Pointer; - - VecBufferPool(const std::string &filename, size_t pool_capacity, size_t block_size); + + VecBufferPool(const std::string &filename, size_t pool_capacity, + size_t block_size); ~VecBufferPool() { close(fd_); } VecBufferPoolHandle get_handle(); - char *acquire_buffer(block_id_t block_id, size_t offset, size_t size, int retry = 0); + char *acquire_buffer(block_id_t block_id, size_t offset, size_t size, + int retry = 0); int get_meta(size_t offset, size_t length, char *buffer); @@ -127,11 +130,10 @@ class VecBufferPool { struct VecBufferPoolHandle { VecBufferPoolHandle(VecBufferPool &pool) : pool(pool), hit_num_(0) {}; VecBufferPoolHandle(VecBufferPoolHandle &&other) - : pool(other.pool), - hit_num_(other.hit_num_) { + : pool(other.pool), hit_num_(other.hit_num_) { other.hit_num_ = 0; } - + ~VecBufferPoolHandle() = default; typedef std::shared_ptr Pointer; diff --git a/src/include/zvec/ailego/buffer/concurrentqueue.h b/src/include/zvec/ailego/buffer/concurrentqueue.h index db4835b1..90edaf97 100644 --- a/src/include/zvec/ailego/buffer/concurrentqueue.h +++ b/src/include/zvec/ailego/buffer/concurrentqueue.h @@ -1,5 +1,5 @@ -// Provides a C++11 implementation of a multi-producer, multi-consumer lock-free queue. -// An overview, including benchmark results, is provided here: +// 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 @@ -8,24 +8,26 @@ // Copyright (c) 2013-2020, 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: +// 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. +// - 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. +// 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. // Also dual-licensed under the Boost Software License (see LICENSE.md) @@ -33,8 +35,8 @@ #if defined(__GNUC__) && !defined(__INTEL_COMPILER) // 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) +// 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" @@ -44,10 +46,11 @@ #endif #if defined(_MSC_VER) && (!defined(_HAS_CXX17) || !_HAS_CXX17) -// VS2019 with /W4 warns about constant conditional expressions but unless /std=c++17 or higher -// does not support `if constexpr`, so we have no choice but to simply disable the warning +// VS2019 with /W4 warns about constant conditional expressions but unless +// /std=c++17 or higher does not support `if constexpr`, so we have no choice +// but to simply disable the warning #pragma warning(push) -#pragma warning(disable: 4127) // conditional expression is constant +#pragma warning(disable : 4127) // conditional expression is constant #endif #if defined(__APPLE__) @@ -57,92 +60,128 @@ #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. +// 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 // Requires C++11. Sorry VS2010. #include #endif -#include // for max_align_t +#include +#include +#include // for CHAR_BIT +#include // for max_align_t #include #include +#include +#include // used for thread exit synchronization +#include // partly for __WINPTHREADS_VERSION if on MinGW-w64 w/ POSIX threading #include -#include #include -#include -#include // for CHAR_BIT -#include -#include // partly for __WINPTHREADS_VERSION if on MinGW-w64 w/ POSIX threading -#include // used for thread exit synchronization - -// 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; } - }; -} } + +// 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; + } +}; +} // namespace details +} // namespace moodycamel #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(); } -} } +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(); +} +} // namespace details +} // namespace moodycamel #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) || defined(__MVS__) || defined(MOODYCAMEL_NO_THREAD_LOCAL) -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; +// 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()); +} +} // namespace details +} // namespace moodycamel +#elif defined(__arm__) || defined(_M_ARM) || defined(__aarch64__) || \ + (defined(__APPLE__) && TARGET_OS_IPHONE) || defined(__MVS__) || \ + defined(MOODYCAMEL_NO_THREAD_LOCAL) +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; + typedef std::size_t thread_id_hash_t; #else - typedef thread_id_numeric_size_t thread_id_hash_t; + typedef thread_id_numeric_size_t thread_id_hash_t; #endif - static thread_id_hash_t prehash(thread_id_t const& x) - { + static thread_id_hash_t prehash(thread_id_t const &x) { #ifndef __APPLE__ - return std::hash()(x); + return std::hash()(x); #else - return *reinterpret_cast(&x); + 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 :-) +// 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) @@ -151,17 +190,25 @@ namespace moodycamel { namespace details { // 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. - inline thread_id_t thread_id() { static MOODYCAMEL_THREADLOCAL int x; return reinterpret_cast(&x); } -} } +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. +inline thread_id_t thread_id() { + static MOODYCAMEL_THREADLOCAL int x; + return reinterpret_cast(&x); +} +} +} #endif // Constexpr if #ifndef MOODYCAMEL_CONSTEXPR_IF -#if (defined(_MSC_VER) && defined(_HAS_CXX17) && _HAS_CXX17) || __cplusplus > 201402L +#if (defined(_MSC_VER) && defined(_HAS_CXX17) && _HAS_CXX17) || \ + __cplusplus > 201402L #define MOODYCAMEL_CONSTEXPR_IF if constexpr #define MOODYCAMEL_MAYBE_UNUSED [[maybe_unused]] #else @@ -172,18 +219,20 @@ namespace moodycamel { namespace details { // Exceptions #ifndef MOODYCAMEL_EXCEPTIONS_ENABLED -#if (defined(_MSC_VER) && defined(_CPPUNWIND)) || (defined(__GNUC__) && defined(__EXCEPTIONS)) || (!defined(_MSC_VER) && !defined(__GNUC__)) +#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_CATCH(...) catch (__VA_ARGS__) #define MOODYCAMEL_RETHROW throw -#define MOODYCAMEL_THROW(expr) throw (expr) +#define MOODYCAMEL_THROW(expr) throw(expr) #else -#define MOODYCAMEL_TRY MOODYCAMEL_CONSTEXPR_IF (true) -#define MOODYCAMEL_CATCH(...) else MOODYCAMEL_CONSTEXPR_IF (false) +#define MOODYCAMEL_TRY MOODYCAMEL_CONSTEXPR_IF(true) +#define MOODYCAMEL_CATCH(...) else MOODYCAMEL_CONSTEXPR_IF(false) #define MOODYCAMEL_RETHROW #define MOODYCAMEL_THROW(expr) #endif @@ -194,15 +243,40 @@ namespace moodycamel { namespace details { #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! +// 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)) +#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)) +#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) @@ -214,18 +288,31 @@ namespace moodycamel { namespace details { #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__) && !defined(__MVS__) -// Assume `thread_local` is fully supported in all other C++11 compilers/platforms -#define MOODYCAMEL_CPP11_THREAD_LOCAL_SUPPORTED // tentatively enabled for now; years ago several users report having problems with it on +// 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__) && !defined(__MVS__) +// Assume `thread_local` is fully supported in all other C++11 +// compilers/platforms +#define MOODYCAMEL_CPP11_THREAD_LOCAL_SUPPORTED // tentatively enabled for now; + // years ago 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. +// 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 @@ -234,54 +321,101 @@ namespace moodycamel { namespace details { #endif #endif -namespace moodycamel { namespace details { +namespace moodycamel { +namespace details { #ifndef MOODYCAMEL_ALIGNAS -// VS2013 doesn't support alignas or alignof, and align() requires a constant literal +// VS2013 doesn't support alignas or alignof, and align() requires a constant +// literal #if defined(_MSC_VER) && _MSC_VER <= 1800 #define MOODYCAMEL_ALIGNAS(alignment) __declspec(align(alignment)) #define MOODYCAMEL_ALIGNOF(obj) __alignof(obj) -#define MOODYCAMEL_ALIGNED_TYPE_LIKE(T, obj) typename details::Vs2013Aligned::value, T>::type - template struct Vs2013Aligned { }; // default, unsupported alignment - template struct Vs2013Aligned<1, T> { typedef __declspec(align(1)) T type; }; - template struct Vs2013Aligned<2, T> { typedef __declspec(align(2)) T type; }; - template struct Vs2013Aligned<4, T> { typedef __declspec(align(4)) T type; }; - template struct Vs2013Aligned<8, T> { typedef __declspec(align(8)) T type; }; - template struct Vs2013Aligned<16, T> { typedef __declspec(align(16)) T type; }; - template struct Vs2013Aligned<32, T> { typedef __declspec(align(32)) T type; }; - template struct Vs2013Aligned<64, T> { typedef __declspec(align(64)) T type; }; - template struct Vs2013Aligned<128, T> { typedef __declspec(align(128)) T type; }; - template struct Vs2013Aligned<256, T> { typedef __declspec(align(256)) T type; }; +#define MOODYCAMEL_ALIGNED_TYPE_LIKE(T, obj) \ + typename details::Vs2013Aligned::value, T>::type +template +struct Vs2013Aligned {}; // default, unsupported alignment +template +struct Vs2013Aligned<1, T> { + typedef __declspec(align(1)) T type; +}; +template +struct Vs2013Aligned<2, T> { + typedef __declspec(align(2)) T type; +}; +template +struct Vs2013Aligned<4, T> { + typedef __declspec(align(4)) T type; +}; +template +struct Vs2013Aligned<8, T> { + typedef __declspec(align(8)) T type; +}; +template +struct Vs2013Aligned<16, T> { + typedef __declspec(align(16)) T type; +}; +template +struct Vs2013Aligned<32, T> { + typedef __declspec(align(32)) T type; +}; +template +struct Vs2013Aligned<64, T> { + typedef __declspec(align(64)) T type; +}; +template +struct Vs2013Aligned<128, T> { + typedef __declspec(align(128)) T type; +}; +template +struct Vs2013Aligned<256, T> { + typedef __declspec(align(256)) T type; +}; #else - template struct identity { typedef T type; }; +template +struct identity { + typedef T type; +}; #define MOODYCAMEL_ALIGNAS(alignment) alignas(alignment) #define MOODYCAMEL_ALIGNOF(obj) alignof(obj) -#define MOODYCAMEL_ALIGNED_TYPE_LIKE(T, obj) alignas(alignof(obj)) typename details::identity::type +#define MOODYCAMEL_ALIGNED_TYPE_LIKE(T, obj) \ + alignas(alignof(obj)) typename details::identity::type #endif #endif -} } +} // namespace details +} // namespace moodycamel -// TSAN can false report races in lock-free code. To enable TSAN to be used from projects that use this one, -// we can apply per-function compile-time suppression. -// See https://clang.llvm.org/docs/ThreadSanitizer.html#has-feature-thread-sanitizer +// TSAN can false report races in lock-free code. To enable TSAN to be used +// from projects that use this one, we can apply per-function compile-time +// suppression. See +// https://clang.llvm.org/docs/ThreadSanitizer.html#has-feature-thread-sanitizer #define MOODYCAMEL_NO_TSAN #if defined(__has_feature) - #if __has_feature(thread_sanitizer) - #undef MOODYCAMEL_NO_TSAN - #define MOODYCAMEL_NO_TSAN __attribute__((no_sanitize("thread"))) - #endif // TSAN -#endif // TSAN +#if __has_feature(thread_sanitizer) +#undef MOODYCAMEL_NO_TSAN +#define MOODYCAMEL_NO_TSAN __attribute__((no_sanitize("thread"))) +#endif // TSAN +#endif // TSAN // Compiler-specific likely/unlikely hints -namespace moodycamel { namespace details { +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); } +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; } +static inline bool(likely)(bool x) { + return x; +} +static inline bool(unlikely)(bool x) { + return x; +} #endif -} } +} // namespace details +} // namespace moodycamel #ifdef MOODYCAMEL_QUEUE_INTERNAL_DEBUG #include "internal/concurrentqueue_internal_debug.h" @@ -289,28 +423,34 @@ namespace moodycamel { namespace details { 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); - }; +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 +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:: +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; -} +// 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; +} // namespace details // Default traits for the ConcurrentQueue. To change some of the // traits without re-implementing all of them, inherit from this @@ -318,95 +458,117 @@ namespace details { // 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; - - // The number of times to spin before sleeping when waiting on a semaphore. - // Recommended values are on the order of 1000-10000 unless the number of - // consumer threads exceeds the number of idle cores (in which case try 0-100). - // Only affects instances of the BlockingConcurrentQueue. - static const int MAX_SEMA_SPINS = 10000; - - // Whether to recycle dynamically-allocated blocks into an internal free list or - // not. If false, only pre-allocated blocks (controlled by the constructor - // arguments) will be recycled, and all others will be `free`d back to the heap. - // Note that blocks consumed by explicit producers are only freed on destruction - // of the queue (not following destruction of the token) regardless of this trait. - static const bool RECYCLE_ALLOCATED_BLOCKS = false; - - +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; + + // The number of times to spin before sleeping when waiting on a semaphore. + // Recommended values are on the order of 1000-10000 unless the number of + // consumer threads exceeds the number of idle cores (in which case try + // 0-100). Only affects instances of the BlockingConcurrentQueue. + static const int MAX_SEMA_SPINS = 10000; + + // Whether to recycle dynamically-allocated blocks into an internal free list + // or not. If false, only pre-allocated blocks (controlled by the constructor + // arguments) will be recycled, and all others will be `free`d back to the + // heap. Note that blocks consumed by explicit producers are only freed on + // destruction of the queue (not following destruction of the token) + // regardless of this trait. + static const bool RECYCLE_ALLOCATED_BLOCKS = false; + + #ifndef MCDBGQ_USE_RELACY - // Memory allocation can be customized if needed. - // malloc should return nullptr on failure, and handle alignment like std::malloc. + // 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); } + // 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); } + 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, $); } + // 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 }; @@ -421,3322 +583,3825 @@ struct ConcurrentQueueDefaultTraits struct ProducerToken; struct ConsumerToken; -template class ConcurrentQueue; -template class BlockingConcurrentQueue; +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) - { - 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))); - // Note: extra parens around rhs of operator<< is MSVC bug: https://developercommunity2.visualstudio.com/t/C4554-triggers-when-both-lhs-and-rhs-is/10034931 - // silencing the bug requires #pragma warning(disable: 4554) around the calling code and has no effect when done here. - } - - 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 = left.load(std::memory_order_relaxed); - left.store(right.load(std::memory_order_relaxed), std::memory_order_relaxed); - right.store(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 { }; +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) { + 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))); + // Note: extra parens around rhs of operator<< is MSVC bug: + // https://developercommunity2.visualstudio.com/t/C4554-triggers-when-both-lhs-and-rhs-is/10034931 + // silencing the bug requires #pragma warning(disable: 4554) around the + // calling code and has no effect when done here. +} + +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 = left.load(std::memory_order_relaxed); + left.store(right.load(std::memory_order_relaxed), std::memory_order_relaxed); + right.store(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 { }; +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; +typedef RelacyThreadExitListener ThreadExitListener; +typedef RelacyThreadExitNotifier ThreadExitNotifier; #else - class ThreadExitNotifier; - - struct ThreadExitListener - { - typedef void (*callback_t)(void*); - callback_t callback; - void* userData; - - ThreadExitListener* next; // reserved for use by the ThreadExitNotifier - ThreadExitNotifier* chain; // reserved for use by the ThreadExitNotifier - }; - - class ThreadExitNotifier - { - public: - static void subscribe(ThreadExitListener* listener) - { - auto& tlsInst = instance(); - std::lock_guard guard(mutex()); - listener->next = tlsInst.tail; - listener->chain = &tlsInst; - tlsInst.tail = listener; - } - - static void unsubscribe(ThreadExitListener* listener) - { - std::lock_guard guard(mutex()); - if (!listener->chain) { - return; // race with ~ThreadExitNotifier - } - auto& tlsInst = *listener->chain; - listener->chain = nullptr; - 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."); - std::lock_guard guard(mutex()); - for (auto ptr = tail; ptr != nullptr; ptr = ptr->next) { - ptr->chain = nullptr; - ptr->callback(ptr->userData); - } - } - - // Thread-local - static inline ThreadExitNotifier& instance() - { - static thread_local ThreadExitNotifier notifier; - return notifier; - } - - static inline std::mutex& mutex() - { - // Must be static because the ThreadExitNotifier could be destroyed while unsubscribe is called - static std::mutex mutex; - return mutex; - } - - 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 }; }; -} +class ThreadExitNotifier; + +struct ThreadExitListener { + typedef void (*callback_t)(void *); + callback_t callback; + void *userData; + + ThreadExitListener *next; // reserved for use by the ThreadExitNotifier + ThreadExitNotifier *chain; // reserved for use by the ThreadExitNotifier +}; +class ThreadExitNotifier { + public: + static void subscribe(ThreadExitListener *listener) { + auto &tlsInst = instance(); + std::lock_guard guard(mutex()); + listener->next = tlsInst.tail; + listener->chain = &tlsInst; + tlsInst.tail = listener; + } -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; + static void unsubscribe(ThreadExitListener *listener) { + std::lock_guard guard(mutex()); + if (!listener->chain) { + return; // race with ~ThreadExitNotifier + } + auto &tlsInst = *listener->chain; + listener->chain = nullptr; + 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."); + std::lock_guard guard(mutex()); + for (auto ptr = tail; ptr != nullptr; ptr = ptr->next) { + ptr->chain = nullptr; + ptr->callback(ptr->userData); + } + } + + // Thread-local + static inline ThreadExitNotifier &instance() { + static thread_local ThreadExitNotifier notifier; + return notifier; + } + + static inline std::mutex &mutex() { + // Must be static because the ThreadExitNotifier could be destroyed while + // unsubscribe is called + static std::mutex mutex; + return mutex; + } + + 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 }; }; +} // namespace details + + +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; + } -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; + 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); +// 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 +#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); + 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 = 32 * 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 - MOODYCAMEL_CONSTEXPR_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) - { - MOODYCAMEL_CONSTEXPR_IF (INITIAL_IMPLICIT_PRODUCER_HASH_SIZE == 0) return false; - else 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) - { - MOODYCAMEL_CONSTEXPR_IF (INITIAL_IMPLICIT_PRODUCER_HASH_SIZE == 0) return false; - else 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) - { - MOODYCAMEL_CONSTEXPR_IF (INITIAL_IMPLICIT_PRODUCER_HASH_SIZE == 0) return false; - else 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) - { - MOODYCAMEL_CONSTEXPR_IF (INITIAL_IMPLICIT_PRODUCER_HASH_SIZE == 0) return false; - else 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) - { - MOODYCAMEL_CONSTEXPR_IF (INITIAL_IMPLICIT_PRODUCER_HASH_SIZE == 0) return false; - else 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) - { - MOODYCAMEL_CONSTEXPR_IF (INITIAL_IMPLICIT_PRODUCER_HASH_SIZE == 0) return false; - else 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 constexpr 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)) { - 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_acq_rel) == 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), dynamicallyAllocated(true) - { -#ifdef MCDBGQ_TRACKMEM - owner = nullptr; -#endif - } - - template - inline bool is_empty() const - { - MOODYCAMEL_CONSTEXPR_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(MOODYCAMEL_MAYBE_UNUSED index_t i) - { - MOODYCAMEL_CONSTEXPR_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_acq_rel); - 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(MOODYCAMEL_MAYBE_UNUSED index_t i, size_t count) - { - MOODYCAMEL_CONSTEXPR_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_acq_rel); - assert(prevVal + count <= BLOCK_SIZE); - return prevVal + count == BLOCK_SIZE; - } - } - - template - inline void set_all_empty() - { - MOODYCAMEL_CONSTEXPR_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() - { - MOODYCAMEL_CONSTEXPR_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: - static_assert(std::alignment_of::value <= sizeof(T), "The queue does not support types with an alignment greater than their size at this time"); - MOODYCAMEL_ALIGNED_TYPE_LIKE(char[sizeof(T) * BLOCK_SIZE], T) elements; - 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; - 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"); + 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 = 32 * 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 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; - 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. - - MOODYCAMEL_CONSTEXPR_IF (allocMode == CannotAlloc) { - return false; - } - else if (!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; - } - - MOODYCAMEL_CONSTEXPR_IF (!MOODYCAMEL_NOEXCEPT_CTOR(T, U, new (static_cast(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); - - MOODYCAMEL_CONSTEXPR_IF (!MOODYCAMEL_NOEXCEPT_CTOR(T, U, new (static_cast(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) / static_cast::type>(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 MOODYCAMEL_NO_TSAN 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) { - MOODYCAMEL_CONSTEXPR_IF (allocMode == CannotAlloc) { - // Failed to allocate, undo changes (but keep injected blocks) - pr_blockIndexFront = originalBlockIndexFront; - pr_blockIndexSlotsUsed = originalBlockIndexSlotsUsed; - this->tailBlock = startBlock == nullptr ? firstAllocatedBlock : startBlock; - return false; - } - else if (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; - } - - MOODYCAMEL_CONSTEXPR_IF (MOODYCAMEL_NOEXCEPT_CTOR(T, decltype(*itemFirst), new (static_cast(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) { - index_t stopIndex = (currentTailIndex & ~static_cast(BLOCK_SIZE - 1)) + static_cast(BLOCK_SIZE); - if (details::circular_less_than(newTailIndex, stopIndex)) { - stopIndex = newTailIndex; - } - MOODYCAMEL_CONSTEXPR_IF (MOODYCAMEL_NOEXCEPT_CTOR(T, decltype(*itemFirst), new (static_cast(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(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; - } - - MOODYCAMEL_CONSTEXPR_IF (!MOODYCAMEL_NOEXCEPT_CTOR(T, decltype(*itemFirst), new (static_cast(nullptr)) T(details::deref_noexcept(itemFirst)))) { - if (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) / static_cast::type>(BLOCK_SIZE)); - auto indexIndex = (localBlockIndexHead + offset) & (localBlockIndex->size - 1); - - // Iterate the blocks and dequeue - auto index = firstIndex; - do { - auto firstIndexInBlock = index; - index_t 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(); - - MOODYCAMEL_CONSTEXPR_IF (!MOODYCAMEL_NOEXCEPT_CTOR(T, U, new (static_cast(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; - - MOODYCAMEL_CONSTEXPR_IF (!MOODYCAMEL_NOEXCEPT_CTOR(T, U, new (static_cast(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; - } - -#ifdef _MSC_VER -#pragma warning(push) -#pragma warning(disable: 4706) // assignment within conditional expression -#endif - 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) { - index_t stopIndex = (currentTailIndex & ~static_cast(BLOCK_SIZE - 1)) + static_cast(BLOCK_SIZE); - if (details::circular_less_than(newTailIndex, stopIndex)) { - stopIndex = newTailIndex; - } - MOODYCAMEL_CONSTEXPR_IF (MOODYCAMEL_NOEXCEPT_CTOR(T, decltype(*itemFirst), new (static_cast(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(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; - } -#ifdef _MSC_VER -#pragma warning(pop) + // 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 - - 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; - index_t 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 - } - size_t 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! - MOODYCAMEL_CONSTEXPR_IF (allocMode == CannotAlloc) { - return false; - } - else if (!new_block_index()) { - return false; - } - else { - 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) / static_cast::type>(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; + } + + // 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_CPP11_THREAD_LOCAL_SUPPORTED - public: - details::ThreadExitListener threadExitListener; - private: -#endif - #ifdef MOODYCAMEL_QUEUE_INTERNAL_DEBUG - public: - ImplicitProducer* nextImplicitProducer; - private: + explicitProducers.store(nullptr, std::memory_order_relaxed); + implicitProducers.store(nullptr, std::memory_order_relaxed); #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 - if (!Traits::RECYCLE_ALLOCATED_BLOCKS && block->dynamicallyAllocated) { - destroy(block); - } - else { - 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; - } - - MOODYCAMEL_CONSTEXPR_IF (canAlloc == CanAlloc) { - return create(); - } - else { - return nullptr; - } - } - + // 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 + MOODYCAMEL_CONSTEXPR_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 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) - { -#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 - return ptr; - } - } - } - - 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() - { - MOODYCAMEL_CONSTEXPR_IF (INITIAL_IMPLICIT_PRODUCER_HASH_SIZE == 0) { - return; - } - else { - 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) - { - MOODYCAMEL_CONSTEXPR_IF (INITIAL_IMPLICIT_PRODUCER_HASH_SIZE == 0) { - return; - } - else { - // 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); - assert(mainHash != nullptr); // silence clang-tidy and MSVC warnings (hash cannot be null) - 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 - 1u; - - 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 - 1u; - auto empty = details::invalid_thread_id; -#ifdef MOODYCAMEL_CPP11_THREAD_LOCAL_SUPPORTED - auto reusable = details::invalid_thread_id2; - if (mainHash->entries[index].key.compare_exchange_strong(empty, id, std::memory_order_seq_cst, std::memory_order_relaxed) || - mainHash->entries[index].key.compare_exchange_strong(reusable, id, std::memory_order_seq_cst, std::memory_order_relaxed)) { -#else - if (mainHash->entries[index].key.compare_exchange_strong(empty, id, std::memory_order_seq_cst, 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)) { - size_t 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 = static_cast(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)) { - auto producer = static_cast(recycle_or_create_producer(false)); - if (producer == nullptr) { - implicitProducerHashCount.fetch_sub(1, std::memory_order_relaxed); - return nullptr; - } - -#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 - 1u; - auto empty = details::invalid_thread_id; -#ifdef MOODYCAMEL_CPP11_THREAD_LOCAL_SUPPORTED - auto reusable = details::invalid_thread_id2; - if (mainHash->entries[index].key.compare_exchange_strong(reusable, id, std::memory_order_seq_cst, std::memory_order_relaxed)) { - implicitProducerHashCount.fetch_sub(1, std::memory_order_relaxed); // already counted as a used slot - mainHash->entries[index].value = producer; - break; - } -#endif - if (mainHash->entries[index].key.compare_exchange_strong(empty, id, std::memory_order_seq_cst, std::memory_order_relaxed)) { - 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 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 - 1u; - probedKey = id; - if (hash->entries[index].key.compare_exchange_strong(probedKey, details::invalid_thread_id2, std::memory_order_seq_cst, std::memory_order_relaxed)) { - 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 void* aligned_malloc(size_t size) - { - MOODYCAMEL_CONSTEXPR_IF (std::alignment_of::value <= std::alignment_of::value) - return (Traits::malloc)(size); - else { - size_t alignment = std::alignment_of::value; - void* raw = (Traits::malloc)(size + alignment - 1 + sizeof(void*)); - if (!raw) - return nullptr; - char* ptr = details::align_for(reinterpret_cast(raw) + sizeof(void*)); - *(reinterpret_cast(ptr) - 1) = raw; - return ptr; - } - } - - template - static inline void aligned_free(void* ptr) - { - MOODYCAMEL_CONSTEXPR_IF (std::alignment_of::value <= std::alignment_of::value) - return (Traits::free)(ptr); - else - (Traits::free)(ptr ? *(reinterpret_cast(ptr) - 1) : nullptr); - } - - template - static inline U* create_array(size_t count) - { - assert(count > 0); - U* p = static_cast(aligned_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(); - } - aligned_free(p); - } - - template - static inline U* create() - { - void* p = aligned_malloc(sizeof(U)); - return p != nullptr ? new (p) U : nullptr; - } - - template - static inline U* create(A1&& a1) - { - void* p = aligned_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(); - aligned_free(p); - } - -private: - std::atomic producerListTail; - std::atomic producerCount; - - std::atomic initialBlockPoolIndex; - Block* initialBlockPool; - size_t initialBlockPoolSize; - -#ifndef 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; + 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 - std::atomic explicitProducers; - std::atomic implicitProducers; + details::swap_relaxed(explicitProducers, other.explicitProducers); + details::swap_relaxed(implicitProducers, other.implicitProducers); #endif -}; + return *this; + } -template -ProducerToken::ProducerToken(ConcurrentQueue& queue) - : producer(queue.recycle_or_create_producer(true)) -{ - if (producer != nullptr) { - producer->token = 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) { + MOODYCAMEL_CONSTEXPR_IF(INITIAL_IMPLICIT_PRODUCER_HASH_SIZE == 0) + return false; + else return inner_enqueue(item); + } -template -ProducerToken::ProducerToken(BlockingConcurrentQueue& queue) - : producer(reinterpret_cast*>(&queue)->recycle_or_create_producer(true)) -{ - if (producer != nullptr) { - producer->token = this; - } -} + // 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) { + MOODYCAMEL_CONSTEXPR_IF(INITIAL_IMPLICIT_PRODUCER_HASH_SIZE == 0) + return false; + else return inner_enqueue(std::move(item)); + } -template -ConsumerToken::ConsumerToken(ConcurrentQueue& queue) - : itemsConsumedFromCurrent(0), currentProducer(nullptr), desiredProducer(nullptr) -{ - initialOffset = queue.nextExplicitConsumerId.fetch_add(1, std::memory_order_release); - lastKnownGlobalOffset = static_cast(-1); -} + // 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); + } -template -ConsumerToken::ConsumerToken(BlockingConcurrentQueue& queue) - : itemsConsumedFromCurrent(0), currentProducer(nullptr), desiredProducer(nullptr) -{ - initialOffset = reinterpret_cast*>(&queue)->nextExplicitConsumerId.fetch_add(1, std::memory_order_release); - lastKnownGlobalOffset = static_cast(-1); -} + // 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)); + } -template -inline void swap(ConcurrentQueue& a, ConcurrentQueue& b) MOODYCAMEL_NOEXCEPT -{ - a.swap(b); -} + // 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) { + MOODYCAMEL_CONSTEXPR_IF(INITIAL_IMPLICIT_PRODUCER_HASH_SIZE == 0) + return false; + else return inner_enqueue_bulk(itemFirst, count); + } -inline void swap(ProducerToken& a, ProducerToken& b) MOODYCAMEL_NOEXCEPT -{ - a.swap(b); -} + // 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); + } -inline void swap(ConsumerToken& a, ConsumerToken& b) MOODYCAMEL_NOEXCEPT -{ - a.swap(b); -} + // 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) { + MOODYCAMEL_CONSTEXPR_IF(INITIAL_IMPLICIT_PRODUCER_HASH_SIZE == 0) + return false; + else return inner_enqueue(item); + } -template -inline void swap(typename ConcurrentQueue::ImplicitProducerKVP& a, typename ConcurrentQueue::ImplicitProducerKVP& b) MOODYCAMEL_NOEXCEPT -{ - a.swap(b); -} + // 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) { + MOODYCAMEL_CONSTEXPR_IF(INITIAL_IMPLICIT_PRODUCER_HASH_SIZE == 0) + return false; + else 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) { + MOODYCAMEL_CONSTEXPR_IF(INITIAL_IMPLICIT_PRODUCER_HASH_SIZE == 0) + return false; + else 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 constexpr 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< + canAlloc>(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)) { + 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_acq_rel) == 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), + dynamicallyAllocated(true) { +#ifdef MCDBGQ_TRACKMEM + owner = nullptr; +#endif + } + + template + inline bool is_empty() const { + MOODYCAMEL_CONSTEXPR_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(MOODYCAMEL_MAYBE_UNUSED index_t i) { + MOODYCAMEL_CONSTEXPR_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_acq_rel); + 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(MOODYCAMEL_MAYBE_UNUSED index_t i, + size_t count) { + MOODYCAMEL_CONSTEXPR_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_acq_rel); + assert(prevVal + count <= BLOCK_SIZE); + return prevVal + count == BLOCK_SIZE; + } + } + + template + inline void set_all_empty() { + MOODYCAMEL_CONSTEXPR_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() { + MOODYCAMEL_CONSTEXPR_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: + static_assert(std::alignment_of::value <= sizeof(T), + "The queue does not support types with an alignment greater " + "than their size at this time"); + MOODYCAMEL_ALIGNED_TYPE_LIKE(char[sizeof(T) * BLOCK_SIZE], T) elements; + + 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; + 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< + explicit_context>()) { + 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; + 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< + explicit_context>()) { + // We can re-use the block ahead of us, it's empty! + this->tailBlock = this->tailBlock->next; + this->tailBlock->ConcurrentQueue::Block::template reset_empty< + explicit_context>(); + + // 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. + + MOODYCAMEL_CONSTEXPR_IF(allocMode == CannotAlloc) { + return false; + } + else if (!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< + explicit_context>(); + if (this->tailBlock == nullptr) { + newBlock->next = newBlock; + } else { + newBlock->next = this->tailBlock->next; + this->tailBlock->next = newBlock; + } + this->tailBlock = newBlock; + ++pr_blockIndexSlotsUsed; + } + + MOODYCAMEL_CONSTEXPR_IF(!MOODYCAMEL_NOEXCEPT_CTOR( + T, U, + new (static_cast(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); + + MOODYCAMEL_CONSTEXPR_IF(!MOODYCAMEL_NOEXCEPT_CTOR( + T, U, + new (static_cast(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) / + static_cast::type>( + 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< + explicit_context>(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 MOODYCAMEL_NO_TSAN 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< + explicit_context>()) { + 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) { + MOODYCAMEL_CONSTEXPR_IF(allocMode == CannotAlloc) { + // Failed to allocate, undo changes (but keep injected blocks) + pr_blockIndexFront = originalBlockIndexFront; + pr_blockIndexSlotsUsed = originalBlockIndexSlotsUsed; + this->tailBlock = + startBlock == nullptr ? firstAllocatedBlock : startBlock; + return false; + } + else if (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< + explicit_context>(); + 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< + explicit_context>(); + if (block == this->tailBlock) { + break; + } + block = block->next; + } + + MOODYCAMEL_CONSTEXPR_IF(MOODYCAMEL_NOEXCEPT_CTOR( + T, decltype(*itemFirst), + new (static_cast(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) { + index_t stopIndex = + (currentTailIndex & ~static_cast(BLOCK_SIZE - 1)) + + static_cast(BLOCK_SIZE); + if (details::circular_less_than(newTailIndex, stopIndex)) { + stopIndex = newTailIndex; + } + MOODYCAMEL_CONSTEXPR_IF(MOODYCAMEL_NOEXCEPT_CTOR( + T, decltype(*itemFirst), + new (static_cast(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(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; + } + + MOODYCAMEL_CONSTEXPR_IF(!MOODYCAMEL_NOEXCEPT_CTOR( + T, decltype(*itemFirst), + new (static_cast(nullptr)) + T(details::deref_noexcept(itemFirst)))) { + if (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) / + static_cast::type>( + BLOCK_SIZE)); + auto indexIndex = + (localBlockIndexHead + offset) & (localBlockIndex->size - 1); + + // Iterate the blocks and dequeue + auto index = firstIndex; + do { + auto firstIndexInBlock = index; + index_t 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< + explicit_context>( + 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< + explicit_context>( + 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(); + + MOODYCAMEL_CONSTEXPR_IF(!MOODYCAMEL_NOEXCEPT_CTOR( + T, U, + new (static_cast(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; + + MOODYCAMEL_CONSTEXPR_IF(!MOODYCAMEL_NOEXCEPT_CTOR( + T, U, + new (static_cast(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< + implicit_context>(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< + implicit_context>(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; + } + +#ifdef _MSC_VER +#pragma warning(push) +#pragma warning(disable : 4706) // assignment within conditional expression +#endif + 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< + allocMode>()) == 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< + implicit_context>(); + 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) { + index_t stopIndex = + (currentTailIndex & ~static_cast(BLOCK_SIZE - 1)) + + static_cast(BLOCK_SIZE); + if (details::circular_less_than(newTailIndex, stopIndex)) { + stopIndex = newTailIndex; + } + MOODYCAMEL_CONSTEXPR_IF(MOODYCAMEL_NOEXCEPT_CTOR( + T, decltype(*itemFirst), + new (static_cast(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(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; + } +#ifdef _MSC_VER +#pragma warning(pop) +#endif + + 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; + index_t 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< + implicit_context>( + 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< + implicit_context>( + 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 + } + size_t 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! + MOODYCAMEL_CONSTEXPR_IF(allocMode == CannotAlloc) { + return false; + } + else if (!new_block_index()) { + return false; + } + else { + 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) / + static_cast::type>(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 + if (!Traits::RECYCLE_ALLOCATED_BLOCKS && block->dynamicallyAllocated) { + destroy(block); + } else { + 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; + } + + MOODYCAMEL_CONSTEXPR_IF(canAlloc == CanAlloc) { + return create(); + } + else { + 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< + explicit_context>() || + 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) { +#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 + return ptr; + } + } + } + + 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() { + MOODYCAMEL_CONSTEXPR_IF(INITIAL_IMPLICIT_PRODUCER_HASH_SIZE == 0) { + return; + } + else { + 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) { + MOODYCAMEL_CONSTEXPR_IF(INITIAL_IMPLICIT_PRODUCER_HASH_SIZE == 0) { + return; + } + else { + // 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); + assert( + mainHash != + nullptr); // silence clang-tidy and MSVC warnings (hash cannot be null) + 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 - 1u; + + 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 - 1u; + auto empty = details::invalid_thread_id; +#ifdef MOODYCAMEL_CPP11_THREAD_LOCAL_SUPPORTED + auto reusable = details::invalid_thread_id2; + if (mainHash->entries[index].key.compare_exchange_strong( + empty, id, std::memory_order_seq_cst, + std::memory_order_relaxed) || + mainHash->entries[index].key.compare_exchange_strong( + reusable, id, std::memory_order_seq_cst, + std::memory_order_relaxed)) { +#else + if (mainHash->entries[index].key.compare_exchange_strong( + empty, id, std::memory_order_seq_cst, + 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)) { + size_t 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 = static_cast(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)) { + auto producer = + static_cast(recycle_or_create_producer(false)); + if (producer == nullptr) { + implicitProducerHashCount.fetch_sub(1, std::memory_order_relaxed); + return nullptr; + } + +#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 - 1u; + auto empty = details::invalid_thread_id; +#ifdef MOODYCAMEL_CPP11_THREAD_LOCAL_SUPPORTED + auto reusable = details::invalid_thread_id2; + if (mainHash->entries[index].key.compare_exchange_strong( + reusable, id, std::memory_order_seq_cst, + std::memory_order_relaxed)) { + implicitProducerHashCount.fetch_sub( + 1, + std::memory_order_relaxed); // already counted as a used slot + mainHash->entries[index].value = producer; + break; + } +#endif + if (mainHash->entries[index].key.compare_exchange_strong( + empty, id, std::memory_order_seq_cst, + std::memory_order_relaxed)) { + 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 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 - 1u; + probedKey = id; + if (hash->entries[index].key.compare_exchange_strong( + probedKey, details::invalid_thread_id2, + std::memory_order_seq_cst, std::memory_order_relaxed)) { + 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 void *aligned_malloc(size_t size) { + MOODYCAMEL_CONSTEXPR_IF(std::alignment_of::value <= + std::alignment_of::value) + return (Traits::malloc)(size); + else { + size_t alignment = std::alignment_of::value; + void *raw = (Traits::malloc)(size + alignment - 1 + sizeof(void *)); + if (!raw) return nullptr; + char *ptr = details::align_for(reinterpret_cast(raw) + + sizeof(void *)); + *(reinterpret_cast(ptr) - 1) = raw; + return ptr; + } + } + + template + static inline void aligned_free(void *ptr) { + MOODYCAMEL_CONSTEXPR_IF(std::alignment_of::value <= + std::alignment_of::value) + return (Traits::free)(ptr); + else(Traits::free)(ptr ? *(reinterpret_cast(ptr) - 1) : nullptr); + } + + template + static inline U *create_array(size_t count) { + assert(count > 0); + U *p = static_cast(aligned_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(); + } + aligned_free(p); + } + + template + static inline U *create() { + void *p = aligned_malloc(sizeof(U)); + return p != nullptr ? new (p) U : nullptr; + } + + template + static inline U *create(A1 &&a1) { + void *p = aligned_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(); + aligned_free(p); + } + + private: + std::atomic producerListTail; + std::atomic producerCount; + + std::atomic initialBlockPoolIndex; + Block *initialBlockPool; + size_t initialBlockPoolSize; + +#ifndef 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 = static_cast(-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 = static_cast(-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); +} + +} // namespace moodycamel #if defined(_MSC_VER) && (!defined(_HAS_CXX17) || !_HAS_CXX17) #pragma warning(pop) From 55e6f1b5bc3c33e1eefe0a0ef7f2662925f1257b Mon Sep 17 00:00:00 2001 From: "yinzefeng.yzf" Date: Tue, 10 Feb 2026 11:29:48 +0800 Subject: [PATCH 06/11] clang format --- src/core/algorithm/hnsw/hnsw_entity.h | 3 +-- .../zvec/ailego/buffer/concurrentqueue.h | 2 +- src/include/zvec/core/framework/index_storage.h | 17 +++++++++++------ 3 files changed, 13 insertions(+), 9 deletions(-) diff --git a/src/core/algorithm/hnsw/hnsw_entity.h b/src/core/algorithm/hnsw/hnsw_entity.h index d2c06c41..70ea3dcc 100644 --- a/src/core/algorithm/hnsw/hnsw_entity.h +++ b/src/core/algorithm/hnsw/hnsw_entity.h @@ -147,8 +147,7 @@ struct Neighbors { Neighbors(uint32_t cnt_in, const node_id_t *data_in) : cnt{cnt_in}, data{data_in} {} - Neighbors(IndexStorage::MemoryBlock &mem_block) - : neighbor_block{mem_block} { + Neighbors(IndexStorage::MemoryBlock &mem_block) : neighbor_block{mem_block} { auto hd = reinterpret_cast(neighbor_block.data()); cnt = hd->neighbor_cnt; data = hd->neighbors; diff --git a/src/include/zvec/ailego/buffer/concurrentqueue.h b/src/include/zvec/ailego/buffer/concurrentqueue.h index 90edaf97..3b587642 100644 --- a/src/include/zvec/ailego/buffer/concurrentqueue.h +++ b/src/include/zvec/ailego/buffer/concurrentqueue.h @@ -1706,7 +1706,7 @@ class ConcurrentQueue { // contention. template // N must inherit FreeListNode or have the same fields // (and initialization of them) - struct FreeList { + struct FreeList { FreeList() : freeListHead(nullptr) {} FreeList(FreeList &&other) : freeListHead(other.freeListHead.load(std::memory_order_relaxed)) { diff --git a/src/include/zvec/core/framework/index_storage.h b/src/include/zvec/core/framework/index_storage.h index 346b8da4..920580fe 100644 --- a/src/include/zvec/core/framework/index_storage.h +++ b/src/include/zvec/core/framework/index_storage.h @@ -37,7 +37,8 @@ class IndexStorage : public IndexModule { }; MemoryBlock() {} - MemoryBlock(ailego::VecBufferPoolHandle* buffer_pool_handle, int block_id, void *data) + MemoryBlock(ailego::VecBufferPoolHandle *buffer_pool_handle, int block_id, + void *data) : type_(MemoryBlockType::MBT_BUFFERPOOL) { buffer_pool_handle_ = buffer_pool_handle; buffer_block_id_ = block_id; @@ -65,7 +66,8 @@ class IndexStorage : public IndexModule { this->reset(std::move(rhs.data_)); break; case MemoryBlockType::MBT_BUFFERPOOL: - this->reset(std::move(rhs.buffer_pool_handle_), std::move(rhs.buffer_block_id_), std::move(rhs.data_)); + this->reset(std::move(rhs.buffer_pool_handle_), + std::move(rhs.buffer_block_id_), std::move(rhs.data_)); break; default: break; @@ -79,7 +81,8 @@ class IndexStorage : public IndexModule { this->reset(rhs.data_); break; case MemoryBlockType::MBT_BUFFERPOOL: - this->reset(rhs.buffer_pool_handle_, rhs.buffer_block_id_, rhs.data_); + this->reset(rhs.buffer_pool_handle_, rhs.buffer_block_id_, + rhs.data_); buffer_pool_handle_->acquire_one(buffer_block_id_); break; default: @@ -96,7 +99,8 @@ class IndexStorage : public IndexModule { this->reset(std::move(rhs.data_)); break; case MemoryBlockType::MBT_BUFFERPOOL: - this->reset(std::move(rhs.buffer_pool_handle_), std::move(rhs.buffer_block_id_), std::move(rhs.data_)); + this->reset(std::move(rhs.buffer_pool_handle_), + std::move(rhs.buffer_block_id_), std::move(rhs.data_)); break; default: break; @@ -124,7 +128,8 @@ class IndexStorage : public IndexModule { return data_; } - void reset(ailego::VecBufferPoolHandle* buffer_pool_handle, int block_id, void *data) { + void reset(ailego::VecBufferPoolHandle *buffer_pool_handle, int block_id, + void *data) { if (type_ == MemoryBlockType::MBT_BUFFERPOOL) { buffer_pool_handle->release_one(buffer_block_id_); } @@ -145,7 +150,7 @@ class IndexStorage : public IndexModule { MemoryBlockType type_{MBT_UNKNOWN}; void *data_{nullptr}; - mutable ailego::VecBufferPoolHandle* buffer_pool_handle_; + mutable ailego::VecBufferPoolHandle *buffer_pool_handle_; int buffer_block_id_{0}; }; From b24d921053dd2597357c524ebb78d65979948a92 Mon Sep 17 00:00:00 2001 From: "yinzefeng.yzf" Date: Tue, 10 Feb 2026 11:45:03 +0800 Subject: [PATCH 07/11] clang format --- .../zvec/ailego/buffer/concurrentqueue.h | 20 +++++++++---------- 1 file changed, 9 insertions(+), 11 deletions(-) diff --git a/src/include/zvec/ailego/buffer/concurrentqueue.h b/src/include/zvec/ailego/buffer/concurrentqueue.h index 3b587642..16f297e8 100644 --- a/src/include/zvec/ailego/buffer/concurrentqueue.h +++ b/src/include/zvec/ailego/buffer/concurrentqueue.h @@ -709,7 +709,7 @@ struct nomove_if { }; template -static inline auto deref_noexcept(It &it) MOODYCAMEL_NOEXCEPT -> decltype(*it) { +static inline auto deref_noexcept(It &it) MOODYCAMEL_NOEXCEPT->decltype(*it) { return *it; } @@ -2833,10 +2833,9 @@ class ConcurrentQueue { // Create the new block pr_blockIndexSize <<= 1; - auto newRawPtr = static_cast( - (Traits::malloc)(sizeof(BlockIndexHeader) + - std::alignment_of::value - 1 + - sizeof(BlockIndexEntry) * pr_blockIndexSize)); + 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; @@ -3556,12 +3555,11 @@ class ConcurrentQueue { 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)); + 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; } From 8916f90025d24f0bada96ac3c3eac9aa1d9efd7a Mon Sep 17 00:00:00 2001 From: "yinzefeng.yzf" Date: Tue, 10 Feb 2026 15:53:54 +0800 Subject: [PATCH 08/11] clang format --- src/include/zvec/ailego/buffer/concurrentqueue.h | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/src/include/zvec/ailego/buffer/concurrentqueue.h b/src/include/zvec/ailego/buffer/concurrentqueue.h index 16f297e8..f7f3d77e 100644 --- a/src/include/zvec/ailego/buffer/concurrentqueue.h +++ b/src/include/zvec/ailego/buffer/concurrentqueue.h @@ -111,8 +111,8 @@ static inline thread_id_t thread_id() { #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); +extern "C" + __declspec(dllimport) unsigned long __stdcall GetCurrentThreadId(void); namespace moodycamel { namespace details { static_assert(sizeof(unsigned long) == sizeof(std::uint32_t), @@ -709,7 +709,7 @@ struct nomove_if { }; template -static inline auto deref_noexcept(It &it) MOODYCAMEL_NOEXCEPT->decltype(*it) { +static inline auto deref_noexcept(It &it) MOODYCAMEL_NOEXCEPT -> decltype(*it) { return *it; } From e3d014ca629bdc7beda6fed6e32d95c0428470e3 Mon Sep 17 00:00:00 2001 From: Zefeng Yin Date: Tue, 10 Feb 2026 22:41:21 +0800 Subject: [PATCH 09/11] fix bugs --- src/ailego/buffer/buffer_pool.cc | 15 ++++++++------- .../algorithm/flat/flat_streamer_context.h | 10 +++++++++- src/core/algorithm/hnsw/hnsw_context.h | 4 ++++ src/core/interface/index.cc | 7 +++++-- src/core/utility/buffer_storage.cc | 18 ++++++------------ src/include/zvec/ailego/buffer/buffer_pool.h | 5 +++-- .../zvec/core/framework/index_storage.h | 2 +- .../index/column/vector_column_indexer_test.cc | 1 - 8 files changed, 36 insertions(+), 26 deletions(-) diff --git a/src/ailego/buffer/buffer_pool.cc b/src/ailego/buffer/buffer_pool.cc index 81ed92bf..bdbf0a03 100644 --- a/src/ailego/buffer/buffer_pool.cc +++ b/src/ailego/buffer/buffer_pool.cc @@ -66,7 +66,7 @@ void LPMap::init(size_t entry_num) { entries_[i].load_count.store(0); entries_[i].buffer = nullptr; } - cache_.init(entry_num); + cache_.init(entry_num * 4); } char *LPMap::acquire_block(block_id_t block_id) { @@ -136,9 +136,7 @@ void LPMap::recycle(moodycamel::ConcurrentQueue &free_buffers) { } } -VecBufferPool::VecBufferPool(const std::string &filename, size_t pool_capacity, - size_t block_size) - : pool_capacity_(pool_capacity) { +VecBufferPool::VecBufferPool(const std::string &filename) { fd_ = open(filename.c_str(), O_RDONLY); if (fd_ < 0) { throw std::runtime_error("Failed to open file: " + filename); @@ -148,9 +146,12 @@ VecBufferPool::VecBufferPool(const std::string &filename, size_t pool_capacity, throw std::runtime_error("Failed to stat file: " + filename); } file_size_ = st.st_size; +} - size_t buffer_num = pool_capacity_ / block_size; - size_t block_num = file_size_ / block_size + 500; +int VecBufferPool::init(size_t pool_capacity, size_t block_size) { + pool_capacity_ = pool_capacity; + size_t buffer_num = pool_capacity_ / block_size + 10; + size_t block_num = file_size_ / block_size + 10; lp_map_.init(block_num); for (size_t i = 0; i < buffer_num; i++) { char *buffer = (char *)aligned_alloc(64, block_size); @@ -160,6 +161,7 @@ VecBufferPool::VecBufferPool(const std::string &filename, size_t pool_capacity, } LOG_DEBUG("Buffer pool num: %zu, entry num: %zu", buffer_num, lp_map_.entry_num()); + return 0; } VecBufferPoolHandle VecBufferPool::get_handle() { @@ -209,7 +211,6 @@ char *VecBufferPool::acquire_buffer(block_id_t block_id, size_t offset, int VecBufferPool::get_meta(size_t offset, size_t length, char *buffer) { ssize_t read_bytes = pread(fd_, buffer, length, offset); if (read_bytes != static_cast(length)) { - LOG_ERROR("Buffer pool failed to read file at offset: %zu", offset); LOG_ERROR("Buffer pool failed to read file at offset: %zu", offset); return -1; } diff --git a/src/core/algorithm/flat/flat_streamer_context.h b/src/core/algorithm/flat/flat_streamer_context.h index 24cfd9e5..22a1106a 100644 --- a/src/core/algorithm/flat/flat_streamer_context.h +++ b/src/core/algorithm/flat/flat_streamer_context.h @@ -190,10 +190,18 @@ class FlatStreamerContext : public IndexStreamer::Context { group_topk_heaps_.clear(); } - void reset() override {} + void reset() override { + for (auto &it : results_) { + it.clear(); + } + for (auto &it : group_results_) { + it.clear(); + } + } //! Reset the context void reset(const FlatStreamer *owner) { + this->reset(); magic_ = owner->magic(); feature_size_ = owner->meta().element_size(); diff --git a/src/core/algorithm/hnsw/hnsw_context.h b/src/core/algorithm/hnsw/hnsw_context.h index 22bcfaad..e776b81a 100644 --- a/src/core/algorithm/hnsw/hnsw_context.h +++ b/src/core/algorithm/hnsw/hnsw_context.h @@ -335,6 +335,7 @@ class HnswContext : public IndexContext { //! Reset context void reset(void) override { + this->clear(); set_filter(nullptr); reset_threshold(); set_fetch_vector(false); @@ -422,6 +423,9 @@ class HnswContext : public IndexContext { for (auto &it : results_) { it.clear(); } + for (auto &it : group_results_) { + it.clear(); + } } uint32_t *mutable_stats_get_neighbors() { diff --git a/src/core/interface/index.cc b/src/core/interface/index.cc index 038f67d4..72005bc9 100644 --- a/src/core/interface/index.cc +++ b/src/core/interface/index.cc @@ -406,8 +406,9 @@ int Index::Search(const VectorData &vector_data, } // dense support refiner, but sparse doesn't + int ret = 0; if (search_param->refiner_param == nullptr) { - return _dense_search(vector_data, search_param, result, context); + ret = _dense_search(vector_data, search_param, result, context); } else { auto &reference_index = search_param->refiner_param->reference_index; if (reference_index == nullptr) { @@ -441,8 +442,10 @@ int Index::Search(const VectorData &vector_data, // TODO: should copy other params? flat_search_param->bf_pks = std::make_shared>(keys); - return reference_index->Search(vector_data, flat_search_param, result); + ret = reference_index->Search(vector_data, flat_search_param, result); } + context->reset(); + return ret; } diff --git a/src/core/utility/buffer_storage.cc b/src/core/utility/buffer_storage.cc index dcdb13d3..1fccbe2e 100644 --- a/src/core/utility/buffer_storage.cc +++ b/src/core/utility/buffer_storage.cc @@ -98,7 +98,7 @@ class BufferStorage : public IndexStorage { } size_t buffer_offset = segment_header_start_offset_ + segment_header_->content_offset + - segment_->meta()->data_index + offset; + segment_->meta()->data_index; *data = owner_->get_buffer(buffer_offset, capacity_, segment_id_) + offset; return len; @@ -114,7 +114,7 @@ class BufferStorage : public IndexStorage { } size_t buffer_offset = segment_header_start_offset_ + segment_header_->content_offset + - segment_->meta()->data_index + offset; + segment_->meta()->data_index; data.reset( owner_->buffer_pool_handle_.get(), segment_id_, owner_->get_buffer(buffer_offset, capacity_, segment_id_) + offset); @@ -177,21 +177,15 @@ class BufferStorage : public IndexStorage { //! Open storage int open(const std::string &path, bool /*create*/) override { - LOG_INFO("open buffer storage 1"); file_name_ = path; - buffer_pool_ = std::make_shared( - path, 20lu * 1024 * 1024 * 1024, 2490368 * 2); + buffer_pool_ = std::make_shared(path); buffer_pool_handle_ = std::make_shared( buffer_pool_->get_handle()); int ret = ParseToMapping(); - LOG_ERROR("segment count: %lu, max_segment_size: %lu", segments_.size(), - max_segment_size_); - for (auto iter = segments_.begin(); iter != segments_.end(); iter++) { - auto seg = this->get(iter->first, 0); - MemoryBlock block; - int len = seg->read(0, block, 1); - LOG_ERROR("segment %s: %d", iter->first.c_str(), len); + if (ret != 0) { + return ret; } + ret = buffer_pool_->init(20lu * 1024 * 1024 * 1024, max_segment_size_); if (ret != 0) { return ret; } diff --git a/src/include/zvec/ailego/buffer/buffer_pool.h b/src/include/zvec/ailego/buffer/buffer_pool.h index f1a0149c..c27065a2 100644 --- a/src/include/zvec/ailego/buffer/buffer_pool.h +++ b/src/include/zvec/ailego/buffer/buffer_pool.h @@ -97,12 +97,13 @@ class VecBufferPool { public: typedef std::shared_ptr Pointer; - VecBufferPool(const std::string &filename, size_t pool_capacity, - size_t block_size); + VecBufferPool(const std::string &filename); ~VecBufferPool() { close(fd_); } + int init(size_t pool_capacity, size_t block_size); + VecBufferPoolHandle get_handle(); char *acquire_buffer(block_id_t block_id, size_t offset, size_t size, diff --git a/src/include/zvec/core/framework/index_storage.h b/src/include/zvec/core/framework/index_storage.h index 920580fe..9173da3e 100644 --- a/src/include/zvec/core/framework/index_storage.h +++ b/src/include/zvec/core/framework/index_storage.h @@ -131,7 +131,7 @@ class IndexStorage : public IndexModule { void reset(ailego::VecBufferPoolHandle *buffer_pool_handle, int block_id, void *data) { if (type_ == MemoryBlockType::MBT_BUFFERPOOL) { - buffer_pool_handle->release_one(buffer_block_id_); + buffer_pool_handle_->release_one(buffer_block_id_); } type_ = MemoryBlockType::MBT_BUFFERPOOL; buffer_pool_handle_ = buffer_pool_handle; diff --git a/tests/db/index/column/vector_column_indexer_test.cc b/tests/db/index/column/vector_column_indexer_test.cc index 483efcde..251e5a18 100644 --- a/tests/db/index/column/vector_column_indexer_test.cc +++ b/tests/db/index/column/vector_column_indexer_test.cc @@ -2160,7 +2160,6 @@ TEST(VectorColumnIndexerTest, Failure) { ASSERT_TRUE(indexer->Flush().ok()); ASSERT_TRUE(indexer->Close().ok()); { - ailego::BufferManager::Instance().init(10 * 1024 * 1024, 1); auto indexer = std::make_shared( index_file_path, FieldSchema("test", DataType::VECTOR_FP32, 3, false, From ed6a3f205e9aaec904316cf5df318dbc073207d3 Mon Sep 17 00:00:00 2001 From: Zefeng Yin Date: Wed, 11 Feb 2026 10:58:59 +0800 Subject: [PATCH 10/11] =?UTF-8?q?fix=20complie=E2=80=9C?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- src/core/utility/buffer_storage.cc | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/src/core/utility/buffer_storage.cc b/src/core/utility/buffer_storage.cc index 1fccbe2e..4db38cb0 100644 --- a/src/core/utility/buffer_storage.cc +++ b/src/core/utility/buffer_storage.cc @@ -13,7 +13,7 @@ // limitations under the License. #include -// #include +#include #include #include #include @@ -476,7 +476,7 @@ class BufferStorage : public IndexStorage { IndexFormat::MetaFooter footer_; std::map segments_{}; std::map id_hash_{}; - size_t max_segment_size_{0}; + uint64_t max_segment_size_{0}; std::unique_ptr segment_buffer_{nullptr}; ailego::VecBufferPool::Pointer buffer_pool_{nullptr}; @@ -487,4 +487,4 @@ class BufferStorage : public IndexStorage { INDEX_FACTORY_REGISTER_STORAGE(BufferStorage); } // namespace core -} // namespace zvec \ No newline at end of file +} // namespace zvec From 95b1c16dafcd417e15f04d5bda276cc1d8431774 Mon Sep 17 00:00:00 2001 From: "yinzefeng.yzf" Date: Wed, 11 Feb 2026 11:31:05 +0800 Subject: [PATCH 11/11] clang format --- src/core/utility/buffer_storage.cc | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/src/core/utility/buffer_storage.cc b/src/core/utility/buffer_storage.cc index 4db38cb0..d339553a 100644 --- a/src/core/utility/buffer_storage.cc +++ b/src/core/utility/buffer_storage.cc @@ -12,8 +12,8 @@ // See the License for the specific language governing permissions and // limitations under the License. -#include #include +#include #include #include #include