RDMA program randomly hangs

问题:

Anyone out there who has done RDMA programming using the RDMA_CM library?

I'm having a hard time finding even simple examples to study. There's an rdma_client & rdma_server example in librdmacm, but it doesn't run in a loop (rping does loop, but it's written using IB verbs directly instead of rdma_cm functions).

I've put together a trivial ping-pong program, but it locks up anywhere after 1 - 100 bounces. I found adding a sleep inside the client makes it work longer before hanging, which indicates a race condition.

The client gets stuck in rdma_get_send_comp() and the server gets stuck in rdma_get_recv_comp().

My limited understanding is that before every rdma_post_send(), you need to issue a rdma_post_recv() that will be coming after the send. Also before every send (except for the 1st client send), you need to wait for a message (rdma_get_recv()) indicating the other side is ready to receive.

What could be wrong?

Server(rdma_cm_id *id)
{
    ibv_wc wc;

    int ret;
    uint8_t recvBuffer[MESSAGE_BUFFER_SIZE],
            sendBuffer[MESSAGE_BUFFER_SIZE];

    ibv_mr *recvMemRegion =  rdma_reg_msgs(id, recvBuffer, MESSAGE_BUFFER_SIZE);
    if (!recvMemRegion)
      throw 0;

    ibv_mr *sendMemRegion = rdma_reg_msgs(id, sendBuffer, MESSAGE_BUFFER_SIZE);
    if (!sendMemRegion)
      throw 0;


    if (ret = rdma_post_recv(id, NULL, recvBuffer, 1, recvMemRegion))
      throw 0;


    if (ret = rdma_accept(id, NULL))
       throw 0;

    do
    {
        if ((ret = rdma_get_recv_comp(id, &wc)) <= 0)
            throw 0;

        if (ret = rdma_post_recv(id, NULL, recvBuffer, 1, recvMemRegion))
          throw 0;

        if (ret = rdma_post_send(id, NULL, sendBuffer, 1, sendMemRegion, 0))
           throw 0;

        printf(".");
        fflush(stdout);

        if ((ret = rdma_get_send_comp(id, &wc)) <= 0)
          throw 0;
    }
    while (true);
}

Client()   // sends the 1st message
{
    // queue-pair parameters are:
    attr.cap.max_send_wr = attr.cap.max_recv_wr = 4;
    attr.cap.max_send_sge = attr.cap.max_recv_sge = 2;
    attr.cap.max_inline_data = 16;
    attr.qp_context = id;
    attr.sq_sig_all = 1;
    attr.qp_type = IBV_QPT_RC;

    <create connection boiler plate>

    uint8_t recvBuffer[MESSAGE_BUFFER_SIZE],
            sendBuffer[MESSAGE_BUFFER_SIZE];

    recvMemRegion = rdma_reg_msgs(id, recvBuffer, MESSAGE_BUFFER_SIZE);
    if (!recvMemRegion)
        throw 0;

    sendMemRegion = rdma_reg_msgs(id, sendBuffer, MESSAGE_BUFFER_SIZE);
    if (!sendMemRegion)
        throw 0;

    if (ret = rdma_connect(id, NULL))
      throw 0;

    do
    {
        if (ret = rdma_post_recv(id, NULL, recvBuffer, 1, recvMemRegion))
          throw 0;

        //usleep(5000);
        if (ret = rdma_post_send(id, NULL, sendBuffer, 1, sendMemRegion, 0))
          throw 0;

        if ((ret = rdma_get_send_comp(id, &wc)) <= 0)
           throw 0;

        if ((ret = rdma_get_recv_comp(id, &wc)) <= 0)
          throw 0;

        printf(".");
        fflush(stdout);
    } 
    while (true);
}

回答1:

Curses! I was bit by a bug in librdmacm-1.0.15-1 (from 2012) that came with SUSE 11. I knew there was nothing wrong with my send/recv sequencing.

I first tried comparing my code with other examples. In one example I saw

while (!ibv_poll_cq(id->send_cq, 1, &wc));

instead of rdma_get_send_comp() and likewise for rdma_get_recv_comp(). I tried replacing those in my example and miraculously, the hanging is gone!

Hmm, maybe rdma_get_send_comp() isn't doing what I'm expecting. I'd better take a look at the code. I got the code for both 1.0.15 and 1.0.18 and what do I see in rdma_verbs.h?

2 very different IB verb sequences:

// 1.0.15
rdma_get_send_comp(struct rdma_cm_id *id, struct ibv_wc *wc)
{
        struct ibv_cq *cq;
        void *context;
        int ret;

        ret = ibv_poll_cq(id->send_cq, 1, wc);
        if (ret)
                goto out;

        ret = ibv_req_notify_cq(id->send_cq, 0);
        if (ret)
                return rdma_seterrno(ret);

        while (!(ret = ibv_poll_cq(id->send_cq, 1, wc))) {
                ret = ibv_get_cq_event(id->send_cq_channel, &cq, &context);
                if (ret)
                        return rdma_seterrno(ret);

                assert(cq == id->send_cq && context == id);
                ibv_ack_cq_events(id->send_cq, 1);
        }
out:
        return (ret < 0) ? rdma_seterrno(ret) : ret;
}

vs
// 1.0.18
rdma_get_send_comp(struct rdma_cm_id *id, struct ibv_wc *wc)
{
        struct ibv_cq *cq;
        void *context;
        int ret;

        do {
                ret = ibv_poll_cq(id->send_cq, 1, wc);
                if (ret)
                        break;

                ret = ibv_req_notify_cq(id->send_cq, 0);
                if (ret)
                        return rdma_seterrno(ret);

                ret = ibv_poll_cq(id->send_cq, 1, wc);
                if (ret)
                        break;

                ret = ibv_get_cq_event(id->send_cq_channel, &cq, &context);
                if (ret)
                        return ret;

                assert(cq == id->send_cq && context == id);
                ibv_ack_cq_events(id->send_cq, 1);
        } while (1);

        return (ret < 0) ? rdma_seterrno(ret) : ret;
}

Can anyone explain why 1.0.18 works while 1.0.15 randomly hangs?