[zeromq-dev] What is the canonical handling of zeromq sockets when fork+exec?

[zmqdev]

* Background

I have a service that starts workers on demand with fork+exec.
The requests arrive over zeromq sockets.

After the fork, before the exec, I close all file descriptors > 2, 
keeping only stdin/out/err. I then exec the requested program.


* Problem

It works. Except that I get some rare core dumps (of the service) with 
the following assertion failure:

	Bad file descriptor (src/epoll.cpp:90)

and the backtrace:

     #0  0xf77f5430 in __kernel_vsyscall ()
     #1  0xf743f1f7 in raise () from /lib/libc.so.6
     #2  0xf7440a33 in abort () from /lib/libc.so.6
     #3  0xf7067134 in zmq::zmq_abort(char const*) () from $LIBS/libzmq.so.5
     #4  0xf7065e6c in zmq::epoll_t::rm_fd(void*) () from $LIBS/libzmq.so.5
     #5  0xf7068823 in zmq::io_object_t::rm_fd(void*) () from 
$LIBS/libzmq.so.5
     #6  0xf70958af in zmq::stream_engine_t::unplug() () from 
$LIBS/libzmq.so.5
     #7  0xf7098711 in 
zmq::stream_engine_t::error(zmq::stream_engine_t::error_reason_t) () 
from $LIBS/libzmq.so.5
     #8  0xf7098867 in zmq::stream_engine_t::timer_event(int) () from 
$LIBS/libzmq.so.5
     #9  0xf707f972 in zmq::poller_base_t::execute_timers() () from 
$LIBS/libzmq.so.5
     #10 0xf7066209 in zmq::epoll_t::loop() () from $LIBS/libzmq.so.5
     #11 0xf7066467 in zmq::epoll_t::worker_routine(void*) () from 
$LIBS/libzmq.so.5
     #12 0xf709d67e in thread_routine () from $LIBS/libzmq.so.5
     #13 0xf7619b2c in start_thread () from /lib/libpthread.so.0
     #14 0xf750808e in clone () from /lib/libc.so.6

This is with zeromq-4.1.4 on RHEL 7.3 x86_64.

So I wonder: is there some interaction between parent and child?


* Documentation

The Guide and the FAQ do not address explicitly the fork+exec point.

The question has been asked several times on the mailing list in various 
forms, without a definitive answer (for dummies like me at least).


* Questions:

Do I need to zmq_close the sockets in the child?
Or is zmq_term in the child enough?
Does closing the file descriptors in the child cause problems in the parent?

What is the correct way to handle this?