chunked_fifo.hh

/*
 * This file is open source software, licensed to you under the terms
 * of the Apache License, Version 2.0 (the "License").  See the NOTICE file
 * distributed with this work for additional information regarding copyright
 * ownership.  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.
 */
/*
 * Copyright (C) 2016 ScyllaDB Ltd.
 */

#pragma once

// An unbounded FIFO queue of objects of type T.
//
// It provides operations to push items in one end of the queue, and pop them
// from the other end of the queue - both operations are guaranteed O(1)
// (not just amortized O(1)). The size() operation is also O(1).
// chunked_fifo also guarantees that the largest contiguous memory allocation
// it does is O(1). The total memory used is, of course, O(N).
//
// How does chunked_fifo differ from std::list<>, circular_buffer<> and
// std::deque()?
//
// std::list<> can also make all the above guarantees, but is inefficient -
// both at run speed (every operation requires an allocation), and in memory
// use. Much more efficient than std::list<> is our circular_buffer<>, which
// allocates a contiguous array to hold the items and only reallocates it,
// exponentially, when the queue grows. On one test of several different
// push/pop scenarios, circular_buffer<> was between 5 and 20 times faster
// than std::list, and also used considerably less memory.
// The problem with circular_buffer<> is that gives up on the last guarantee
// we made above: circular_buffer<> allocates all the items in one large
// contiguous allocation - that might not be possible when the memory is
// highly fragmented.
// std::deque<> aims to solve the contiguous allocation problem by allocating
// smaller chunks of the queue, and keeping a list of them in an array. This
// array is necessary to allow for O(1) random access to any element, a
// feature which we do not need; But this array is itself contiguous so
// std::deque<> attempts larger contiguous allocations the larger the queue
// gets: std::deque<>'s contiguous allocation is still O(N) and in fact
// exactly 1/64 of the size of circular_buffer<>'s contiguous allocation.
// So it's an improvement over circular_buffer<>, but not a full solution.
//
// chunked_fifo<> is such a solution: it also allocates the queue in fixed-
// size chunks (just like std::deque) but holds them in a linked list, not
// a contiguous array, so there are no large contiguous allocations.
//
// Unlike std::deque<> or circular_buffer<>, chunked_fifo only provides the
// operations needed by std::queue, i.e.,: empty(), size(), front(), back(),
// push_back() and pop_front(). For simplicity, we do *not* implement other
// possible operations, like inserting or deleting elements from the "wrong"
// side of the queue or from the middle, nor random-access to items in the
// middle of the queue or iteration over the items without popping them.
//
// Another feature of chunked_fifo which std::deque is missing is the ability
// to control the chunk size, as a template parameter. In std::deque the
// chunk size is undocumented and fixed - in gcc, it is always 512 bytes.
// chunked_fifo, on the other hand, makes the chunk size (in number of items
// instead of bytes) a template parameter; In situations where the queue is
// expected to become very long, using a larger chunk size might make sense
// because it will result in fewer allocations.
//
// chunked_fifo uses uninitialized storage for unoccupied elements, and thus
// uses move/copy constructors instead of move/copy assignments, which are
// less efficient.

#include <memory>
#include <algorithm>

template <typename T, size_t items_per_chunk = 128>
class chunked_fifo {
    static_assert((items_per_chunk & (items_per_chunk - 1)) == 0,
            "chunked_fifo chunk size must be power of two");
    // Note: a "chunk" struct should be allocated but not constructed, so that
    // the individual items are not constructed until first pushed.
    struct chunk {
        T items[items_per_chunk];
        struct chunk* next;
        // begin and end interpreted mod items_per_chunk
        unsigned begin;
        unsigned end;
    };
    // We pop from the chunk at _front_chunk. This chunk is then linked to
    // the following chunks via the "next" link. _back_chunk points to the
    // last chunk in this list, and it is where we push.
    chunk* _front_chunk = nullptr; // where we pop
    chunk* _back_chunk = nullptr; // where we push
    // We want an O(1) size but don't want to maintain a size() counter
    // because this will slow down every push and pop operation just for
    // the rare size() call. Instead, we just keep a count of chunks (which
    // doesn't change on every push or pop), from which we can calculate
    // size() when needed, and still be O(1).
    // This assumes the invariant that all middle chunks (except the front
    // and back) are always full.
    size_t _nchunks = 0;
    // A list of freed chunks, to support reserve() and to improve
    // performance of repeated push and pop, especially on an empty queue.
    // It is a performance/memory tradeoff how many freed chunks to keep
    // here (see save_free_chunks constant below).
    chunk* _free_chunks = nullptr;
    size_t _nfree_chunks = 0;
public:
    using value_type = T;
    using size_type = size_t;
    using reference = T&;
    using pointer = T*;
    using const_reference = const T&;
    using const_pointer = const T*;
public:
    chunked_fifo() = default;
    chunked_fifo(chunked_fifo&& x) noexcept;
    chunked_fifo(const chunked_fifo& X) = delete;
    ~chunked_fifo();
    chunked_fifo& operator=(const chunked_fifo&) = delete;
    chunked_fifo& operator=(chunked_fifo&&) noexcept;
    inline void push_back(const T& data);
    inline void push_back(T&& data);
    T& back();
    const T& back() const;
    template <typename... A>
    inline void emplace_back(A&&... args);
    inline T& front() const noexcept;
    inline void pop_front() noexcept;
    inline bool empty() const noexcept;
    inline size_t size() const noexcept;
    void clear() noexcept;
    // reserve(n) ensures that at least (n - size()) further push() calls can
    // be served without needing new memory allocation.
    // Calling pop()s between these push()es is also allowed and does not
    // alter this guarantee.
    // Note that reserve() does not reduce the amount of memory already
    // reserved - use shrink_to_fit() for that.
    void reserve(size_t n);
    // shrink_to_fit() frees memory held, but unused, by the queue. Such
    // unused memory might exist after pops, or because of reserve().
    void shrink_to_fit();
private:
    void back_chunk_new();
    void front_chunk_delete() noexcept;
    inline void ensure_room_back();
    void undo_room_back();
    inline size_t mask(size_t idx) const noexcept;

};

template <typename T, size_t items_per_chunk>
inline
chunked_fifo<T, items_per_chunk>::chunked_fifo(chunked_fifo&& x) noexcept
        : _front_chunk(x._front_chunk)
        , _back_chunk(x._back_chunk)
        , _nchunks(x._nchunks)
        , _free_chunks(x._free_chunks)
        , _nfree_chunks(x._nfree_chunks) {
    x._front_chunk = nullptr;
    x._back_chunk = nullptr;
    x._nchunks = 0;
    x._free_chunks = nullptr;
    x._nfree_chunks = 0;
}

template <typename T, size_t items_per_chunk>
inline
chunked_fifo<T, items_per_chunk>&
chunked_fifo<T, items_per_chunk>::operator=(chunked_fifo&& x) noexcept {
    if (&x != this) {
        this->~chunked_fifo();
        new (this) chunked_fifo(std::move(x));
    }
    return *this;
}

template <typename T, size_t items_per_chunk>
inline size_t
chunked_fifo<T, items_per_chunk>::mask(size_t idx) const noexcept {
    return idx & (items_per_chunk - 1);
}

template <typename T, size_t items_per_chunk>
inline bool
chunked_fifo<T, items_per_chunk>::empty() const noexcept {
    return _front_chunk == nullptr;
}

template <typename T, size_t items_per_chunk>
inline size_t
chunked_fifo<T, items_per_chunk>::size() const noexcept{
    if (_front_chunk == nullptr) {
        return 0;
    } else if (_back_chunk == _front_chunk) {
        // Single chunk.
        return _front_chunk->end - _front_chunk->begin;
    } else {
        return _front_chunk->end - _front_chunk->begin
                +_back_chunk->end - _back_chunk->begin
                + (_nchunks - 2) * items_per_chunk;
    }
}

template <typename T, size_t items_per_chunk>
void chunked_fifo<T, items_per_chunk>::clear() noexcept {
#if 1
    while (!empty()) {
        pop_front();
    }
#else
    // This is specialized code to free the contents of all the chunks and the
    // chunks themselves. but since destroying a very full queue is not an
    // important use case to optimize, the simple loop above is preferable.
    if (!_front_chunk) {
        // Empty, nothing to do
        return;
    }
    // Delete front chunk (partially filled)
    for (auto i = _front_chunk->begin; i != _front_chunk->end; ++i) {
        _front_chunk->items[mask(i)].~T();
    }
    chunk *p = _front_chunk->next;
    operator delete(_front_chunk, sizeof(chunk));
    // Delete all the middle chunks (all completely filled)
    if (p) {
        while (p != _back_chunk) {
            // These are full chunks
            chunk *nextp = p->next;
            for (auto i = 0; i != items_per_chunk; ++i) {
                // Note we delete out of order (we don't start with p->begin).
                // That should be fine..
                p->items[i].~T();
        }
            operator delete(p, sizeof(chunk));
            p = nextp;
        }
        // Finally delete back chunk (partially filled)
        for (auto i = _back_chunk->begin; i != _back_chunk->end; ++i) {
            _back_chunk->items[mask(i)].~T();
        }
        operator delete(_back_chunk, sizeof(chunk));
    }
    _front_chunk = nullptr;
    _back_chunk = nullptr;
    _nchunks = 0;
#endif
}

template <typename T, size_t items_per_chunk> void
chunked_fifo<T, items_per_chunk>::shrink_to_fit() {
    while (_free_chunks) {
        auto next = _free_chunks->next;
        operator delete(_free_chunks, sizeof(chunk));
        _free_chunks = next;
    }
    _nfree_chunks = 0;
}

template <typename T, size_t items_per_chunk>
chunked_fifo<T, items_per_chunk>::~chunked_fifo() {
    clear();
    shrink_to_fit();
}

template <typename T, size_t items_per_chunk>
void
chunked_fifo<T, items_per_chunk>::back_chunk_new() {
    chunk *old = _back_chunk;
    if (_free_chunks) {
        _back_chunk = _free_chunks;
        _free_chunks = _free_chunks->next;
        --_nfree_chunks;
    } else {
        _back_chunk = static_cast<chunk*>(operator new(sizeof(chunk)));
    }
    _back_chunk->next = nullptr;
    _back_chunk->begin = 0;
    _back_chunk->end = 0;
    if (old) {
        old->next = _back_chunk;
    }
    if (_front_chunk == nullptr) {
        _front_chunk = _back_chunk;
    }
    _nchunks++;
}


template <typename T, size_t items_per_chunk>
inline void
chunked_fifo<T, items_per_chunk>::ensure_room_back() {
    // If we don't have a back chunk or it's full, we need to create a new one
    if (_back_chunk == nullptr ||
            (_back_chunk->end - _back_chunk->begin) == items_per_chunk) {
        back_chunk_new();
    }
}

template <typename T, size_t items_per_chunk>
void
chunked_fifo<T, items_per_chunk>::undo_room_back() {
    // If we failed creating a new item after ensure_room_back() created a
    // new empty chunk, we must remove it, or empty() will be incorrect
    // (either immediately, if the fifo was empty, or when all the items are
    // popped, if it already had items).
    if (_back_chunk->begin == _back_chunk->end) {
        operator delete(_back_chunk, sizeof(chunk));
        --_nchunks;
        if (_nchunks == 0) {
            _back_chunk = nullptr;
            _front_chunk = nullptr;
        } else {
            // Because we don't usually pop from the back, we don't have a "prev"
            // pointer so we need to find the previous chunk the hard and slow
            // way. B
            chunk *old = _back_chunk;
            _back_chunk = _front_chunk;
            while (_back_chunk->next != old) {
                _back_chunk = _back_chunk->next;
            }
            _back_chunk->next = nullptr;
        }
    }

}

template <typename T, size_t items_per_chunk>
template <typename... Args>
inline void
chunked_fifo<T, items_per_chunk>::emplace_back(Args&&... args) {
    ensure_room_back();
    auto p = &_back_chunk->items[mask(_back_chunk->end)];
    try {
        new(p) T(std::forward<Args>(args)...);
    } catch(...) {
        undo_room_back();
        throw;
    }
    ++_back_chunk->end;
}

template <typename T, size_t items_per_chunk>
inline void
chunked_fifo<T, items_per_chunk>::push_back(const T& data) {
    ensure_room_back();
    auto p = &_back_chunk->items[mask(_back_chunk->end)];
    try {
        new(p) T(data);
    } catch(...) {
        undo_room_back();
        throw;
    }
    ++_back_chunk->end;
}

template <typename T, size_t items_per_chunk>
inline void
chunked_fifo<T, items_per_chunk>::push_back(T&& data) {
    ensure_room_back();
    auto p = &_back_chunk->items[mask(_back_chunk->end)];
    try {
        new(p) T(std::move(data));
    } catch(...) {
        undo_room_back();
        throw;
    }
    ++_back_chunk->end;
}

template <typename T, size_t items_per_chunk>
inline
T&
chunked_fifo<T, items_per_chunk>::back() {
    return _back_chunk->items[mask(_back_chunk->end - 1)];
}

template <typename T, size_t items_per_chunk>
inline
const T&
chunked_fifo<T, items_per_chunk>::back() const {
    return _back_chunk->items[mask(_back_chunk->end - 1)];
}

template <typename T, size_t items_per_chunk>
inline T&
chunked_fifo<T, items_per_chunk>::front() const noexcept {
    return _front_chunk->items[mask(_front_chunk->begin)];
}

template <typename T, size_t items_per_chunk>
inline void
chunked_fifo<T, items_per_chunk>::front_chunk_delete() noexcept {
    chunk *next = _front_chunk->next;
    // Certain use cases may need to repeatedly allocate and free a chunk -
    // an obvious example is an empty queue to which we push, and then pop,
    // repeatedly. Another example is pushing and popping to a non-empty queue
    // we push and pop at different chunks so we need to free and allocate a
    // chunk every items_per_chunk operations.
    // The solution is to keep a list of freed chunks instead of freeing them
    // immediately. There is a performance/memory tradeoff of how many freed
    // chunks to save: If we save them all, the queue can never shrink from
    // its maximum memory use (this is how circular_buffer behaves).
    // The ad-hoc choice made here is to limit the number of saved chunks to 1,
    // but this could easily be made a configuration option.
    static constexpr int save_free_chunks = 1;
    if (_nfree_chunks < save_free_chunks) {
        _front_chunk->next = _free_chunks;
        _free_chunks = _front_chunk;
        ++_nfree_chunks;
    } else {
        operator delete(_front_chunk, sizeof(chunk));
    }
    // If we only had one chunk, _back_chunk is gone too.
    if (_back_chunk == _front_chunk) {
        _back_chunk = nullptr;
    }
    _front_chunk = next;
    --_nchunks;
}

template <typename T, size_t items_per_chunk>
inline void
chunked_fifo<T, items_per_chunk>::pop_front() noexcept {
    front().~T();
    // If the front chunk has become empty, we need to free remove it and use
    // the next one.
    if (++_front_chunk->begin == _front_chunk->end) {
        front_chunk_delete();
    }
}

template <typename T, size_t items_per_chunk>
void chunked_fifo<T, items_per_chunk>::reserve(size_t n) {
    // reserve() guarantees that (n - size()) additional push()es will
    // succeed without reallocation:
    size_t need = n - size();
    // If we already have a back chunk, it might have room for some pushes
    // before filling up, so decrease "need":
    if (_back_chunk) {
        need -= items_per_chunk - (_back_chunk->end - _back_chunk->begin);
    }
    size_t needed_chunks = (need + items_per_chunk - 1) / items_per_chunk;
    // If we already have some freed chunks saved, we need to allocate fewer
    // additional chunks, or none at all
    if (needed_chunks <= _nfree_chunks) {
        return;
    }
    needed_chunks -= _nfree_chunks;
    while (needed_chunks--) {
        chunk *c = static_cast<chunk*>(operator new(sizeof(chunk)));
        c->next = _free_chunks;
        _free_chunks = c;
        ++_nfree_chunks;
    }
}