在链表 mpi rma 中插入新节点时出错

Question

我正在尝试在 mpi RMA 中实现无锁分布式链表。

这里是插入函数的代码，在给定键的节点后插入一个新节点。

void insertAfter(int id, int newVal, int key, int rank, nodePtr head, MPI_Win win)
{
    nodePtr newNode, curNodePtr = head, next, tmp2 = { 0 }, fetched = { 0 }, tmp3 = { 0 };
    node tmp, curNode = { 0 }, test;
    int done = 0;
    
    newNode.rank = rank;
    newNode.disp = allocElem(id, newVal, win);

    while (done == 0 && curNodePtr.rank != -1 && curNodePtr.disp != (MPI_Aint)MPI_BOTTOM) {
        MPI_Get_accumulate(NULL, 0, MPI_BYTE, (void*)&curNode, sizeof(node),
            MPI_BYTE, curNodePtr.rank, curNodePtr.disp, sizeof(node), MPI_BYTE, MPI_NO_OP, win);

        MPI_Win_flush(curNodePtr.rank, win);

        if (curNode.id == key) {
            next = curNode.next;

            do{
                MPI_Compare_and_swap((void*)&newNode.rank, (void*)&next.rank, (void*)&tmp2.rank,
                    MPI_INT, curNodePtr.rank, (MPI_Aint) & (((node*)curNodePtr.disp)->next.rank), win);

                MPI_Win_flush(curNodePtr.rank, win);

                if (tmp2.rank == next.rank) {

                    MPI_Accumulate((void*) & next.rank, 1, MPI_INT, newNode.rank,
                        (MPI_Aint) & (((node*)newNode.disp)->next.rank), 1, MPI_INT,
                        MPI_REPLACE, win);

                    MPI_Accumulate((void*) & next.disp, 1, MPI_AINT, newNode.rank,
                        (MPI_Aint) & (((node*)newNode.disp)->next.disp), 1, MPI_AINT,
                        MPI_REPLACE, win);

                    MPI_Win_flush(newNode.rank, win);
                
                    MPI_Accumulate(&newNode.disp, 1, MPI_AINT, curNodePtr.rank,
                        (MPI_Aint) & (((node*)curNodePtr.disp)->next.disp), 1, MPI_AINT,
                        MPI_REPLACE, win);

                    MPI_Win_flush(curNodePtr.rank, win);

                    done = 1;
                }
                else {
                    do {
                        MPI_Fetch_and_op((void*)&fetched.disp, (void*)&tmp2.disp, MPI_AINT, curNodePtr.rank, 
                            (MPI_Aint) & (((node*)curNodePtr.disp)->next.disp), MPI_NO_OP, win);
                        MPI_Win_flush(curNodePtr.rank, win);
                    } while (tmp2.disp == next.disp);
                }
                next = tmp2;
            } while (done != 1);
        }
        curNodePtr = curNode.next;
    }
}

这个函数在列表中找到正确的位置并尝试添加一个新节点。当我使用 mpiexec -n 1 启动程序时，一切正常，但是当我使用 -n > 2 启动它时，它不起作用。它说“请求的位移指定了 RMA 窗口之外的内存”。当我使用 -n = 2 启动它时它工作正常，但是当我使用此功能打印列表时：

void printList(int procid, nodePtr head, MPI_Win win)
{
    nodePtr curNodePtr = head;
    printf("Rank[%d]: Result list is: \n", procid);

    while (curNodePtr.disp != nullPtr.disp) {

        node curNode = { 0 };

        MPI_Win_lock(MPI_LOCK_SHARED, curNodePtr.rank, 0, win);

        MPI_Get(&curNode, sizeof(node), MPI_BYTE,
            curNodePtr.rank, curNodePtr.disp, sizeof(node), MPI_BYTE, win);

        printf("id %d: val %d was inserted by rank %d at displacement %x marked %d nextR %d nextD %x"
            "\n", curNode.id, curNode.val, curNodePtr.rank, curNodePtr.disp, curNode.marked,
            curNode.next.rank, curNode.next.disp);

        MPI_Win_unlock(curNodePtr.rank, win);

        curNodePtr = curNode.next;
    }
}

它表明 nextR = nextD = 0 在由等级为 1 的进程添加的节点上（由等级 0 插入的节点具有有效的 nextR 和 nextD 值）

这里是 allocElem 的代码

MPI_Aint allocElem(int id, int val, MPI_Win win) {
    MPI_Aint disp;
    node* allocNode;

    MPI_Alloc_mem(sizeof(node), MPI_INFO_NULL, &allocNode);

    allocNode->id = id;
    allocNode->val = val;
    allocNode->next = nullPtr;
    allocNode->marked = 0;

    MPI_Win_attach(win, allocNode, sizeof(node));

    if (allocNodeCount == allocNodeSize) {
        allocNodeSize += 100;
        allocNodesTmp = (node**)realloc(allocNodes, allocNodeSize * sizeof(node*));
        if (allocNodesTmp != NULL)
            allocNodes = allocNodesTmp;
        else {
            printf("Error while allocating memory!\n");
            return -1;
        }
    }

    allocNodes[allocNodeCount] = allocNode;
    allocNodeCount++;

    MPI_Get_address(allocNode, &disp);

    return disp;
}

这是主要功能：

        MPI_Win_lock_all(0, win);

    for (int i = 0; i < 10; i++) {
        insertAfter(2,3, -1, procid, head, win);
    }

    MPI_Win_unlock_all(win);

    MPI_Barrier(MPI_COMM_WORLD);

    if (procid == 0) printList(procid, head, win);

    MPI_Barrier(MPI_COMM_WORLD);

    MPI_Win_free(&win);

    for (int i = 0; i < allocNodeCount; i++) MPI_Free_mem(allocNodes[i]);

    MPI_Finalize();

启动 n = 2 的程序给出下一个输出