[zeromq-dev] zmq_poll: timeout issue

[Ilja Golshtein]

Martin, Matt,

thank you very much for clarification
and for speedy answering.

To be honest I don't fill why spurious wakeups are unavoidable in 0mq ... though it doesn't matter.

Here is a mockup of client application

==
    //  One I/O thread in the thread pool will do.
    zmq::context_t ctx (1);
    std::auto_ptr<zmq::socket_t> sp;
    sp.reset(new zmq::socket_t(ctx, ZMQ_REQ/*ZMQ_XREQ*/));

    //  Connect to the server.
    sp->connect ("tcp://localhost:23001");

    for (int i = 0; i != 20; i++)
    {
      zmq::message_t request (10);
      strcpy((char*)request.data (), "AAAAAAAAA");
      sp->send (request);
      std::cout << "sent" << std::endl;
        

      zmq_pollitem_t items[1];
      items[0].socket = *sp.get();;
      items[0].events = ZMQ_POLLIN;
          
      struct timespec tp1;
      ::clock_gettime(CLOCK_MONOTONIC, &tp1);
      for (;;)
      {
        int poll_ret = zmq::poll(items, 1, 3000000);
        if (items[0].revents == ZMQ_POLLIN)
        {
          assert(poll_ret > 0);
            
          //  Get the reply. 
          zmq::message_t reply;
          sp->recv(&reply);
          std::cout << (char*)reply.data() << std::endl;
          break;
        }
        else
        {
          std::cout << "no event fired" <<
            "poll_ret=" << poll_ret << std::endl;
          struct timespec tp2;
          ::clock_gettime(CLOCK_MONOTONIC, &tp2);
          if (tp2.tv_sec - tp1.tv_sec < 3)
          {
            continue;
          }
          else
          {
            std::cout << "reconnecting - message lost" << std::endl;
            sp.reset(new zmq::socket_t(ctx, ZMQ_REQ));
            sp->connect ("tcp://localhost:23001");
            break;
          }
        }
      } 
    }
==

The goals are
1. Don't got stuck if response is lost
2. Not block if server unavailable.

The only reason of recreating socket is REQ/REP paradigm (perhaps XREQ/XREP are more suitable).

Suggestions are appreciated.

Thanks.


10.08.2010, 18:07, "Matt Weinstein" <matt_weinstein at yahoo.com>:
> This is the same problem as pthread condition variables have, e.g.:
>
> http://www.justsoftwaresolutions.co.uk/threading/condition-variable-spurious-wakes.html
>
> Just think of the revents as condition variables to check...
>
> Does someone want to write a quick wrapper ?
>
> PS I recommend CLOCK_MONOTONIC, but it may not be available, and
> requires -lrt ...
>
--
Best regards
Ilja Golshtein