source: doc/theses/thierry_delisle_PhD/code/readyQ_proto/links2.hpp @ fcd65ca

#pragma once

#include <assert.h>

#include "utils.hpp"

//------------------------------------------------------------
// Queue based on the MCS lock
// It is a Multi-Producer/Single-Consumer queue threads pushing
// elements must hold on to the elements they push
// Not appropriate for an async message queue for example,
template<typename node_t>
class mcs_queue {
        node_t * volatile tail;

public:
        mcs_queue(): tail(nullptr) {}

        inline bool empty() const { return !tail; }

        node_t * push( node_t * elem ) {
                /* paranoid */ assert(!elem->_links.next);
                // Race to add to the tail
                node_t * prev = __atomic_exchange_n(&tail, elem, __ATOMIC_SEQ_CST);
                // If we aren't the first, we need to tell the person before us
                // No need to
                if (prev) prev->_links.next = elem;
                return prev;
        }

        // Advances the head of the list, dropping the element given.
        // Passing an element that is not the head is undefined behavior
        // NOT Multi-Thread Safe, concurrent pushes are safe
        node_t * advance(node_t * elem) {
                node_t * expected = elem;
                // Check if this is already the last item
                if (__atomic_compare_exchange_n(&tail, &expected, nullptr, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST)) return nullptr;

                // If not wait for next item to show-up, filled by push
                while (!elem->_links.next) Pause();

                // we need to return if the next link was empty
                node_t * ret = elem->_links.next;

                // invalidate link to reset to initial state
                elem->_links.next = nullptr;
                return ret;
        }
};

//------------------------------------------------------------
// Queue based on the MCS lock
// Extension of the above lock which supports 'blind' pops.
// i.e., popping a value from the head without knowing what the head is
// has no extra guarantees beyond the mcs_queue
template<typename node_t>
class mpsc_queue : private mcs_queue<node_t> {
        node_t * volatile _head;
public:
        mpsc_queue(): mcs_queue<node_t>(), _head(nullptr) {}

        inline bool empty() const { return mcs_queue<node_t>::empty(); }

        node_t * head() const { return _head; }

        // Added a new element to the queue
        // Multi-Thread Safe, Lock-Free
        inline node_t * push(node_t * elem) {
                node_t * prev = mcs_queue<node_t>::push(elem);
                if (!prev) _head = elem;
                return prev;
        }

        // Pop an element from the queue
        // return the element that was removed
        // next is set to the new head of the queue
        // NOT Multi-Thread Safe
        inline node_t * pop(node_t *& next) {
                node_t * elem = _head;
                // If head is empty just return
                if (!elem) return nullptr;

                // If there is already someone in the list, then it's easy
                if (elem->_links.next) {
                        _head = next = elem->_links.next;
                        // force memory sync
                        __atomic_thread_fence(__ATOMIC_SEQ_CST);

                        // invalidate link to reset to initial state
                        elem->_links.next = nullptr;
                }
                // Otherwise, there might be a race where it only looks but someone is enqueuing
                else {
                        // null out head here, because we linearize with push
                        // at the CAS in advance and therefore can write to head
                        // after that point, it could overwrite the write in push
                        _head = nullptr;
                        next = mcs_queue<node_t>::advance(elem);

                        // Only write to the head if there is a next element
                        // it is the only way we can guarantee we are not overwriting
                        // a write made in push
                        if (next) _head = next;
                }

                // return removed element
                return elem;
        }

        // Same as previous function
        inline node_t * pop() {
                node_t * _ = nullptr;
                return pop(_);
        }
};