buffer_queue.hpp

Go to the documentation of this file.

// -----------------------------------------------------------------------------------------------------
// Copyright (c) 2006-2023, Knut Reinert & Freie Universität Berlin
// Copyright (c) 2016-2023, Knut Reinert & MPI für molekulare Genetik
// This file may be used, modified and/or redistributed under the terms of the 3-clause BSD-License
// shipped with this file and also available at: https://github.com/seqan/seqan3/blob/master/LICENSE.md
// -----------------------------------------------------------------------------------------------------
 
#pragma once
 
#include <algorithm>
#include <atomic>
#include <bit>
#include <cassert>
#include <cmath>
#include <concepts>
#include <mutex>
#include <ranges>
#include <seqan3/std/new>
#include <shared_mutex>
#include <span>
#include <type_traits>
#include <vector>
 
#include <seqan3/core/range/type_traits.hpp>
#include <seqan3/utility/container/concept.hpp>
#include <seqan3/utility/parallel/detail/spin_delay.hpp>
 
namespace seqan3::contrib
{
 
enum class queue_op_status : uint8_t
{
    success = 0,
    empty,
    full,
    closed
};
 
enum struct buffer_queue_policy : uint8_t
{
    fixed,
    dynamic
};
 
// Ringbuffer implementation:
// The underlying buffer has size (number of actual elements + 1). This is a trick to easily check if the queue is empty
// or full. Furthermore, the ring buffer uses 4 pointers. The actual push_back and pop_front position as well as
// the pending push_back and pop_front position. The latter indicates the position that have been advanced concurrently
// by multiple threads from either end.
//
// head: position to read/extract from the queue (first inserted elment) => pop_front_position
// tail: position where to write to/push new elements to the queue => push_back_position
// head_read: The actual position after x threads concurrently popped from the queue. => pending_pop_front_position
// tail_write: The actual position after x threads concurrently pushed to the queue. => pending_push_back_position
//  [  ?  ]  [  4  ]  [  3  ]  [  8  ]  [  0  ]  [  x  ]  [  ?  ]
//                       |                          ^
//                       v                          |
//             head            headRead   tail  tailWrite
//
// valid buffer between  [headRead, tail)
// currently filled      [tail, tailWrite)
// currently removed     [head, headRead)
//
// State: empty = (head == tail)
//  [  ?  ]  [  ?  ]  [  ?  ]  [  ?  ]  [  ?  ]  [  ?  ]  [  ?  ]
//                                        tail
//                                        head
// The head is on the same position as tail.
// This means that currently no element is in the buffer.
 
// State: full = (tail + 1 == head)
//  [  2  ]  [  4  ]  [  3  ]  [  ?  ]  [  8  ]  [  0  ]  [  7  ]
//                               tail
//                                        head
// The tail is one position before the head.
// This means that currently no element can be added to the buffer since it is full.
// Strategies are to either wait until some elements have been popped or to expand the capacity of the
// queue by one, inserting the element at the current tail position and moving all elements starting from head one
// position to the right.
 
template <std::semiregular value_t,
          sequence_container buffer_t = std::vector<value_t>,
          buffer_queue_policy buffer_policy = buffer_queue_policy::dynamic>
class buffer_queue
{
public:
    using buffer_type = buffer_t;
    using value_type = typename buffer_type::value_type;
    using size_type = typename buffer_type::size_type;
    using reference = void;
    using const_reference = void;
 
    // Default constructor sets capacity to 1 (still empty)
    buffer_queue() : buffer_queue{0u}
    {}
    buffer_queue(buffer_queue const &) = delete;
    buffer_queue(buffer_queue &&) = delete;
    buffer_queue & operator=(buffer_queue const &) = delete;
    buffer_queue & operator=(buffer_queue &&) = delete;
    ~buffer_queue() = default;
 
    // you can set the initial capacity here
    explicit buffer_queue(size_type const init_capacity)
    {
        buffer.resize(init_capacity + 1);
        ring_buffer_capacity = std::bit_ceil(buffer.size());
    }
 
    template <std::ranges::input_range range_type>
        requires std::convertible_to<std::ranges::range_value_t<range_type>, value_type>
    buffer_queue(size_type const init_capacity, range_type && r) : buffer_queue{init_capacity}
    {
        std::ranges::copy(r, std::ranges::begin(buffer));
    }
 
    template <typename value2_t>
        requires std::convertible_to<value2_t, value_t>
    void push(value2_t && value)
    {
        detail::spin_delay delay{};
 
        for (;;)
        {
            auto status = try_push(std::forward<value2_t>(value));
            if (status == queue_op_status::closed)
                throw queue_op_status::closed;
            else if (status == queue_op_status::success)
                return;
 
            assert(status != queue_op_status::empty);
            assert(status == queue_op_status::full);
            delay.wait(); // pause and then try again.
        }
    } // throws if closed
 
    template <typename value2_t>
        requires std::convertible_to<value2_t, value_t>
    queue_op_status wait_push(value2_t && value)
    {
        detail::spin_delay delay{};
 
        for (;;)
        {
            auto status = try_push(std::forward<value2_t>(value));
            // wait until queue is not full anymore..
            if (status != queue_op_status::full)
                return status;
 
            assert(status != queue_op_status::empty);
            assert(status == queue_op_status::full);
            delay.wait(); // pause and then try again.
        }
    }
 
    value_type value_pop() // throws if closed
    {
        detail::spin_delay delay{};
 
        value_type value{};
        for (;;)
        {
            if (!writer_waiting.load())
            {
                auto status = try_pop(value);
 
                if (status == queue_op_status::closed)
                    throw queue_op_status::closed;
                else if (status == queue_op_status::success)
                    return value;
 
                assert(status != queue_op_status::full);
                assert(status == queue_op_status::empty);
            }
            delay.wait(); // pause and then try again.
        }
    }
 
    queue_op_status wait_pop(value_type & value)
    {
        detail::spin_delay delay{};
 
        queue_op_status status;
        for (;;)
        {
            if (!writer_waiting.load())
            {
                status = try_pop(value);
 
                if (status == queue_op_status::closed || status == queue_op_status::success)
                    break;
 
                assert(status != queue_op_status::full);
                assert(status == queue_op_status::empty);
            }
            delay.wait(); // pause and then try again.
        }
        return status;
    }
 
    template <typename value2_t>
        requires std::convertible_to<value2_t, value_t>
    queue_op_status try_push(value2_t &&);
 
    queue_op_status try_pop(value_t &);
 
    void close()
    {
        if (writer_waiting.exchange(true)) // First writer that closes the queue will continue, the rest returns.
            return;
 
        try
        {
            std::unique_lock write_lock{mutex};
            closed_flag = true;
            writer_waiting.store(false); // reset the lock.
        }
        catch (...)
        {
            writer_waiting.store(false); // reset the lock.
            std::rethrow_exception(std::current_exception());
        }
    }
 
    bool is_closed() const noexcept
    {
        return closed_flag;
    }
 
    bool is_empty() const noexcept
    {
        std::unique_lock write_lock(mutex);
        return pop_front_position == push_back_position;
    }
 
    bool is_full() const noexcept
    {
        std::unique_lock write_lock(mutex);
        return is_ring_buffer_exhausted(pop_front_position, push_back_position);
    }
 
    size_type size() const noexcept
    {
        std::unique_lock write_lock(mutex);
        if (to_buffer_position(pop_front_position) <= to_buffer_position(push_back_position))
        {
            return to_buffer_position(push_back_position) - to_buffer_position(pop_front_position);
        }
        else
        {
            assert(buffer.size() > (to_buffer_position(pop_front_position) - to_buffer_position(push_back_position)));
            return buffer.size() - (to_buffer_position(pop_front_position) - to_buffer_position(push_back_position));
        }
    }
 
private:
    constexpr bool is_ring_buffer_exhausted(size_type const from, size_type const to) const
    {
        assert(to <= (from + ring_buffer_capacity + 1)); // The tail cannot overwrite the head.
 
        return to >= from + ring_buffer_capacity;
    }
 
    constexpr size_type to_buffer_position(size_type const position) const
    {
        return position & (ring_buffer_capacity - 1);
    }
 
    size_type cyclic_increment(size_type position)
    {
        // invariants:
        //   - ring_buffer_capacity is a power of 2
        //   - (position % ring_buffer_capacity) is in [0, buffer.size())
        //
        // return the next greater position that fulfils the invariants
        if (to_buffer_position(++position) >= buffer.size())
            position += ring_buffer_capacity - buffer.size(); // If the position reached
        return position;
    }
 
    template <typename value2_t>
        requires (std::convertible_to<value2_t, value_t>) && (buffer_policy == buffer_queue_policy::fixed) bool
    overflow(value2_t &&)
    {
        return false;
    }
 
    template <typename value2_t>
        requires (std::convertible_to<value2_t, value_t>) && (buffer_policy == buffer_queue_policy::dynamic) bool
    overflow(value2_t && value);
 
    buffer_t buffer;
    alignas(std::hardware_destructive_interference_size) std::shared_mutex mutable mutex{};
    alignas(std::hardware_destructive_interference_size) std::atomic<size_type> pop_front_position{0};
    alignas(std::hardware_destructive_interference_size) std::atomic<size_type> pending_pop_front_position{0};
    alignas(std::hardware_destructive_interference_size) std::atomic<size_type> push_back_position{0};
    alignas(std::hardware_destructive_interference_size) std::atomic<size_type> pending_push_back_position{0};
    alignas(std::hardware_destructive_interference_size) std::atomic<size_type> ring_buffer_capacity{0};
    alignas(std::hardware_destructive_interference_size) std::atomic_bool writer_waiting{false};
    alignas(std::hardware_destructive_interference_size) bool closed_flag{false};
};
 
// Specifies a fixed size buffer queue.
template <std::semiregular value_t, sequence_container buffer_t = std::vector<value_t>>
using fixed_buffer_queue = buffer_queue<value_t, buffer_t, buffer_queue_policy::fixed>;
 
// Specifies a dynamic size buffer queue (growable).
template <std::semiregular value_t, sequence_container buffer_t = std::vector<value_t>>
using dynamic_buffer_queue = buffer_queue<value_t, buffer_t, buffer_queue_policy::dynamic>;
 
// ============================================================================
// Metafunctions
// ============================================================================
 
// ============================================================================
// Functions
// ============================================================================
 
template <std::semiregular value_t, sequence_container buffer_t, buffer_queue_policy buffer_policy>
template <typename value2_t>
    requires (std::convertible_to<value2_t, value_t>) && (buffer_policy == buffer_queue_policy::dynamic)
inline bool buffer_queue<value_t, buffer_t, buffer_policy>::overflow(value2_t && value)
{
    // try to extend capacity
    std::unique_lock write_lock{mutex};
 
    size_type old_size = buffer.size();
    size_type ring_buffer_capacity = this->ring_buffer_capacity;
    size_type local_front = this->pop_front_position;
    size_type local_back = this->push_back_position;
 
    // Expects no pending pushes or pops in unique lock.
    assert(local_back == this->pending_push_back_position);
    assert(local_front == this->pending_pop_front_position);
 
    bool valueWasAppended = false;
 
    // did we reach the capacity limit (another thread could have done the upgrade already)?
    // buffer is full if tail_pos + 1 == head_pos
    if (is_ring_buffer_exhausted(local_front, cyclic_increment(local_back)))
    {
        // In case of a full queue write the value into the additional slot.
        // Note, that the ring-buffer implementation uses one additional field which is not used except
        // when overflow happens. This invariant is used, to simply check for the full/empty state of the queue.
        if (old_size != 0)
        {
            auto it = std::ranges::begin(buffer) + to_buffer_position(local_back);
            *it = std::forward<value2_t>(value);
            local_back = local_front + ring_buffer_capacity;
            valueWasAppended = true;
        }
 
        assert(is_ring_buffer_exhausted(local_front, local_back));
 
        // get positions of head/tail in current buffer sequence
        size_type front_buffer_position = to_buffer_position(local_front);
        size_type back_buffer_position = to_buffer_position(local_back);
 
        // increase capacity by one and move all elements from current pop_front_position one to the right.
        buffer.resize(old_size + 1);
        ring_buffer_capacity = std::bit_ceil(buffer.size());
        std::ranges::move_backward(std::span{buffer.data() + front_buffer_position, buffer.data() + old_size},
                                   buffer.data() + buffer.size());
 
        // Update the pop_front and push_back positions.
        if (old_size != 0)
        {
            this->pending_pop_front_position = this->pop_front_position = front_buffer_position + 1;
            this->pending_push_back_position = this->push_back_position = back_buffer_position + ring_buffer_capacity;
        }
        this->ring_buffer_capacity = ring_buffer_capacity;
    }
    return valueWasAppended;
}
 
// ----------------------------------------------------------------------------
// Function try_pop()
// ----------------------------------------------------------------------------
 
/*
 * @fn ConcurrentQueue#tryPopFront
 * @headerfile <seqan/parallel.h>
 * @brief Try to dequeue a value from a queue.
 *
 * @signature bool tryPopFront(result, queue[, parallelTag]);
 *
 *
 * @param[in,out] queue       A queue.
 * @param[out]    result      The dequeued value (if available).
 * @param[in]     parallelTag The concurrency scheme. If multiple threads dequeue values concurrently this tag must be
 *                            @link ParallelismTags#Parallel @endlink. The more efficient @link ParallelismTags#Serial
 *                            @endlink tag can only be used if one thread calls <tt>popFront</tt> at a time.
 *                            Default is @link ParallelismTags#Parallel @endlink.
 * @return        bool        Returns <tt>true</tt> if a value could be dequeued and <tt>false</tt> otherwise.
 */
template <std::semiregular value_t, sequence_container buffer_t, buffer_queue_policy buffer_policy>
inline queue_op_status buffer_queue<value_t, buffer_t, buffer_policy>::try_pop(value_t & result)
{
    // try to extract a value
    std::shared_lock read_lock{mutex};
 
    size_type local_pending_pop_front_position{};
    size_type next_local_pop_front_position{};
    detail::spin_delay spinDelay{};
 
    local_pending_pop_front_position = this->pending_pop_front_position;
    // wait for queue to become filled
    while (true)
    {
        size_type local_push_back_position = this->push_back_position;
 
        assert(local_pending_pop_front_position <= local_push_back_position);
 
        // Check if queue is empty
        if (local_pending_pop_front_position == local_push_back_position)
        {
            return is_closed() ? queue_op_status::closed : queue_op_status::empty;
        }
 
        // Get the next ring-buffer position to read from.
        next_local_pop_front_position = cyclic_increment(local_pending_pop_front_position);
        // Did another/other thread(s) already acquired this slot?
        // If yes, try with next position. If not, break and read from aquired position.
        if (this->pending_pop_front_position.compare_exchange_weak(local_pending_pop_front_position,
                                                                   next_local_pop_front_position))
            break;
 
        spinDelay.wait();
    }
 
    // Store the value from the aquired read position.
    result = std::ranges::iter_move(buffer.begin() + to_buffer_position(local_pending_pop_front_position));
 
    // wait for pending previous reads and synchronize pop_front_position to local_pending_pop_front_position
    {
        detail::spin_delay delay{};
        size_type acquired_slot = local_pending_pop_front_position;
        while (!this->pop_front_position.compare_exchange_weak(acquired_slot, next_local_pop_front_position))
        {
            acquired_slot = local_pending_pop_front_position;
            delay.wait(); // add adapting delay in case of high contention.
        }
    }
 
    return queue_op_status::success;
}
 
// ----------------------------------------------------------------------------
// Function try_push()
// ----------------------------------------------------------------------------
 
/*
 * @fn ConcurrentQueue#appendValue
 * @headerfile <seqan/parallel.h>
 * @brief Enqueue a value to a queue.
 *
 * @signature void appendValue(queue, val[, expandTag[, parallelTag]);
 *
 *
 * @param[in,out] queue       A queue.
 * @param[in]     val         The value to enqueue.
 * @param[in]     expandTag   The overflow strategy. If @link OverflowStrategyTags#Generous @endlink the queue will be
 *                            automatically resized if the capacity is exceeded, otherwise the thread spinlocks until
 *                            the element can be enqueued.
 *                            Default is the @link DefaultOverflowImplicit @endlink result for the <tt>queue</tt> type.
 * @param[in]     parallelTag The concurrency scheme. If multiple threads enqueue values concurrently this tag must be
 *                            @link ParallelismTags#Parallel @endlink. The more efficient @link ParallelismTags#Serial
 *                            @endlink tag can only be used if one thread calls <tt>appendValue</tt> at a time.
 *                            Default is @link ParallelismTags#Parallel @endlink.
 */
template <std::semiregular value_t, sequence_container buffer_t, buffer_queue_policy buffer_policy>
template <typename value2_t>
    requires std::convertible_to<value2_t, value_t>
inline queue_op_status buffer_queue<value_t, buffer_t, buffer_policy>::try_push(value2_t && value)
{
    // try to push the value
    {
        detail::spin_delay delay{};
 
        std::shared_lock read_lock(mutex);
 
        if (is_closed())
            return queue_op_status::closed;
 
        // Current up to date position to push an element to
        size_type local_pending_push_back_position = this->pending_push_back_position;
 
        while (true)
        {
            // Get the next potential position to write the value too.
            size_type next_local_push_back_position = cyclic_increment(local_pending_push_back_position);
            size_type local_pop_front_position = this->pop_front_position;
 
            // Check if there are enough slots to write to.
            // If not either wait or try to overflow if it is a dynamic queue.
            if (is_ring_buffer_exhausted(local_pop_front_position, next_local_push_back_position))
                break;
 
            // Did another/other thread(s) acquired the current pending position before this thread
            // If yes, try again if not, write into acquired slot.
            if (this->pending_push_back_position.compare_exchange_weak(local_pending_push_back_position,
                                                                       next_local_push_back_position))
            {
                // Current thread acquired the local_pending_push_back_position and can now write the value into the
                // proper slot of the ring buffer.
                auto it = std::ranges::begin(buffer) + to_buffer_position(local_pending_push_back_position);
                *it = std::forward<value2_t>(value);
 
                // wait for pending previous writes and synchronise push_back_position to
                // local_pending_push_back_position
                {
                    detail::spin_delay delay{};
                    // the slot this thread acquired to write to
                    size_type acquired_slot = local_pending_push_back_position;
                    while (
                        !this->push_back_position.compare_exchange_weak(acquired_slot, next_local_push_back_position))
                    {
                        acquired_slot = local_pending_push_back_position;
                        delay.wait();
                    }
                }
                return queue_op_status::success;
            }
 
            delay.wait();
        }
    }
 
    // if possible extend capacity and return.
    if (overflow(std::forward<value2_t>(value)))
    {
        return queue_op_status::success; // always return success, since the queue resizes and cannot be full.
    }
 
    // We could not extend the queue so it must be full.
    return queue_op_status::full;
}
} // namespace seqan3::contrib