tcp/socks_connecters do not handle ZMQ_RECONNECT_IVL of -1

[ehilscher]

Issue description

The ZMQ socket option ZMQ_RECONNECT_IVL supports a value of -1, which is intended to turn off reconnection attempts. When an initial connection attempt fails, tcp_connecter and socks_connecter use options.reconnect_ivl to set a timer to retry the initial connection. But they do not check for -1, and so the timer value is invalid (observed to be a point in the past). It could be argued that a timer should not be set at all when options.reconnect_ivl is -1.

Possible solution:
zmq::tcp_connecter_t::add_reconnect_timer and zmq::socks_connecter_t::add_reconnect_timer are modified:
if (options.reconnect_ivl != -1) {
const int interval = get_new_reconnect_ivl();
add_timer(interval, reconnect_timer_id);
socket->event_connect_retried(endpoint, interval);
reconnect_timer_started = true;
}

An alternative solution is to add an additional socket option (such as ZMQ_CONNECTION_RETRY_IVL) that is only used by the connecters, and could also be -1. This would allow users to let ZMQ retry the initial connection, but not a dropped connection, if desired.

Environment

• libzmq version: 4.2.3
• OS: Windows 10

Steps to reproduce the issue

1. Place a breakpoint in zmq::tcp_connecter_t::add_reconnect_timer.
2. Create a socket.
3. Set the ZMQ_RECONNECT_IVL to -1
4. Set the ZMQ_RECONNECT_IVL_MAX to 0 (should be the default)
5. Call connect on the socket with a TCP address (ensure that there is no listening server at that address).
6. Step through zmq::tcp_connecter_t::add_reconnect_timer:
   Psuedocode:
   int interval = -1 + generate_random () % -1;
   uint64_t expiration = clock.now_ms () + interval; // With MSVC this always results in some time in the past

What's the actual result?

A timer is set for some time in the past to retry an initial connection.

What's the expected result?

A timer is not set to retry an initial connection.

[bluca]

I can't quite reproduce this. With wireshark it's very easy to see - a connection to a non-existing peer is attempted only once with IVL at -1, but many many times without it.
What is the problem exactly?

[ehilscher]

I am seeing the opposite effect. Here is my entire test program:

void* context = zmq_ctx_new();
void* socket = zmq_socket(context, ZMQ_REQ);

int ivl = -1;
zmq_setsockopt(socket, ZMQ_RECONNECT_IVL, &ivl, sizeof(ivl));

zmq_connect(socket, "tcp://127.0.0.1:59999");

while(true)
{
	std::this_thread::sleep_for(std::chrono::milliseconds(1000));
}

I've attached a screenshot of the packet capture from letting this program run for about 30 seconds (filtered by port 59999).

[bluca]

Very weird. Maybe something to do with Windows.
With the fix you suggested, does it work? If so please send a PR

[ehilscher]

Created PR 3163.

My theory is that even on other platforms this problem occurs, but it's partially dependent on the generated random number, and how the % operator handles -1. So even on your system I would bet a reconnect timer is getting set, but it may not trigger for a day, or a month, or even 100 years.

[bluca]

Fixed by #3163