source: libcfa/src/containers/lockfree.hfa @ be5f0a5

#pragma once

#include <assert.h>

#include <stdint.h>
#include <bits/defs.hfa>

forall( T &) {
        //------------------------------------------------------------
        // 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,
        struct mcs_queue {
                T * volatile tail;
        };

        static inline void ?{}(mcs_queue(T) & this) { this.tail = 0p; }
        static inline bool empty(const mcs_queue(T) & this) { return !this.tail; }

        static inline forall(| { T * volatile & ?`next ( T * ); })
        {
                // Adds an element to the list
                // Multi-Thread Safe, Lock-Free
                T * push(mcs_queue(T) & this, T * elem) __attribute__((artificial));
                T * push(mcs_queue(T) & this, T * elem) {
                        /* paranoid */ verify(!(elem`next));
                        // Race to add to the tail
                        T * prev = __atomic_exchange_n(&this.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`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
                T * advance(mcs_queue(T) & this, T * elem) __attribute__((artificial));
                T * advance(mcs_queue(T) & this, T * elem) {
                        T * expected = elem;
                        // Check if this is already the last item
                        if (__atomic_compare_exchange_n(&this.tail, &expected, 0p, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST)) return 0p;

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

                        // we need to return if the next link was empty
                        T * ret = elem`next;

                        // invalidate link to reset to initial state
                        elem`next = 0p;
                        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
        struct mpsc_queue {
                inline mcs_queue(T);
                T * volatile head;
        };

        static inline void ?{}(mpsc_queue(T) & this) {
                ((mcs_queue(T)&)this){};
                this.head = 0p;
        }

        static inline forall(| { T * volatile & ?`next ( T * ); })
        {
                // Added a new element to the queue
                // Multi-Thread Safe, Lock-Free
                T * push(mpsc_queue(T) & this, T * elem) __attribute__((artificial));
                T * push(mpsc_queue(T) & this, T * elem) {
                        T * prev = push((mcs_queue(T)&)this, elem);
                        if (!prev) this.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
                T * pop(mpsc_queue(T) & this, T *& next) __attribute__((artificial));
                T * pop(mpsc_queue(T) & this, T *& next) {
                        T * elem = this.head;
                        // If head is empty just return
                        if (!elem) return 0p;

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

                                // invalidate link to reset to initial state
                                elem`next = 0p;
                        }
                        // 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
                                this.head = 0p;
                                next = advance((mcs_queue(T)&)this, 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) this.head = next;
                        }

                        // return removed element
                        return elem;
                }

                // Same as previous function
                T * pop(mpsc_queue(T) & this) {
                        T * _ = 0p;
                        return pop(this, _);
                }
        }

        //------------------------------------------------------------
        // Queue based on the MCS lock with poisoning
        // 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
        // poisoning the queue prevents any new elements from being push
        // enum(void*) poison_state {
        //      EMPTY = 0p,
        //      POISON = 1p,
        //      IN_PROGRESS = 1p
        // };

        struct poison_list {
                T * volatile head;
        };

        static inline void ?{}(poison_list(T) & this) { this.head = 0p; }

        static inline forall(| { T * volatile & ?`next ( T * ); })
        {
                // Adds an element to the list
                // Multi-Thread Safe, Lock-Free
                T * push(poison_list(T) & this, T * elem) __attribute__((artificial));
                T * push(poison_list(T) & this, T * elem) {
                        /* paranoid */ verify(0p == (elem`next));
                        __atomic_store_n( &elem`next, (T*)1p, __ATOMIC_RELAXED );

                        // read the head up-front
                        T * expected = this.head;
                        for() {
                                // check if it's poisoned
                                if(expected == 1p) return 0p;

                                // try to CAS the elem in
                                if(__atomic_compare_exchange_n(&this.head, &expected, elem, true, __ATOMIC_SEQ_CST, __ATOMIC_RELAXED)) {
                                        // We managed to exchange in, we are done

                                        // We should never succeed the CAS if it's poisonned.
                                        /* paranoid */ verify( expected != 1p );

                                        // If we aren't the first, we need to tell the person before us
                                        // No need to
                                        elem`next = expected;
                                        return expected;
                                }
                        }
                }

                // 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
                T * advance(T * elem) __attribute__((artificial));
                T * advance(T * elem) {
                        T * ret;

                        // Wait for next item to show-up, filled by push
                        while (1p == (ret = __atomic_load_n(&elem`next, __ATOMIC_RELAXED))) Pause();

                        return ret;
                }

                // Poison the queue, preveting new pushes and returning the head
                T * poison(poison_list(T) & this) __attribute__((artificial));
                T * poison(poison_list(T) & this) {
                        T * ret = __atomic_exchange_n( &this.head, (T*)1p, __ATOMIC_SEQ_CST );
                        /* paranoid */ verify( ret != (T*)1p );
                        return ret;
                }
        }
}

forall( T & )
union Link {
        struct {                                                                                        // 32/64-bit x 2
                T * volatile top;                                                               // pointer to stack top
                uintptr_t count;                                                                // count each push
        };
        #if __SIZEOF_INT128__ == 16
        __int128                                                                                        // gcc, 128-bit integer
        #else
        uint64_t                                                                                        // 64-bit integer
        #endif // __SIZEOF_INT128__ == 16
        atom;
}; // Link

forall( T | sized(T) | { Link(T) * ?`next( T * ); } ) {
        struct StackLF {
                Link(T) stack;
        }; // StackLF

        static inline {
                void ?{}( StackLF(T) & this ) with(this) { stack.atom = 0; }

                T * top( StackLF(T) & this ) with(this) { return stack.top; }

                void push( StackLF(T) & this, T & n ) with(this) {
                        *( &n )`next = stack;                                           // atomic assignment unnecessary, or use CAA
                        for () {                                                                        // busy wait
                          if ( __atomic_compare_exchange_n( &stack.atom, &( &n )`next->atom, (Link(T))@{ {&n, ( &n )`next->count + 1} }.atom, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST ) ) break; // attempt to update top node
                        } // for
                } // push

                T * pop( StackLF(T) & this ) with(this) {
                        Link(T) t @= stack;                                                     // atomic assignment unnecessary, or use CAA
                        for () {                                                                        // busy wait
                          if ( t.top == 0p ) return 0p;                         // empty stack ?
                          if ( __atomic_compare_exchange_n( &stack.atom, &t.atom, (Link(T))@{ {( t.top )`next->top, t.count} }.atom, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST ) ) return t.top; // attempt to update top node
                        } // for
                } // pop

                bool unsafe_remove( StackLF(T) & this, T * node ) with(this) {
                        Link(T) * link = &stack;
                        for() {
                                T * next = link->top;
                                if( next == node ) {
                                        link->top = ( node )`next->top;
                                        return true;
                                }
                                if( next == 0p ) return false;
                                link = ( next )`next;
                        }
                }
        } // distribution
} // distribution