Embedding ZeroMQ in the libev Event Loop

In a previous article on ZeroMQ we went over how ZeroMQ is triggered when you use the socket that ZeroMQ returns, in that article there was some discussion of embedding ZeroMQ into another event loop. Let's do that.

libev is an absolutely fantastic library that helps make it easy to write evented programs. Evented programs work by getting notified that an action has happened, and acting upon it. Unlike threaded where multiple pieces of work are being executed at the same time, in an evented system you move every item that could block to an event loop, that then calls back into user code with a notification to continue. If one event uses up more than its fair share of CPU time because it is busy doing a long calculation, every single other event that is waiting will never get notified.

Now, as previously discussed ZeroMQ is edge triggered, so embedding it into an event loop that is level triggered doesn't do us much good, because we will miss certain ZeroMQ notifications.

One way to solve this problem is by looping over ZeroMQ's event system until we get back a notification that it no longer has anything else for us to process, that would look something like this¹:

int zevents = 0;
size_t zevents_len = sizeof(zevents);
zmq_socket.getsockopt(ZMQ_EVENTS, &zevents, &zevents_len);

do {
    if (zevents & ZMQ_POLLIN) {
        // We can read from the ZeroMQ socket
    } else {
        break;
    }

    // Check to see if there is more to read ...
    zmq_socket.getsockopt(ZMQ_EVENTS, &zevents, &zevents_len);
} while (zevents & ZMQ_POLLIN);

if (zevents & ZMQ_POLLOUT) {
    // We can write to the ZeroMQ socket
}

// If neither of the above is true, then it was a false positive

However if we are receiving information from ZeroMQ remote endpoints faster than we can process them, we end up being stuck in that do ... while loop forever. If we have other events we want to process, that isn't entirely fair since they will never ever get called again. Especially in a server application where it may be servicing thousands of clients this is simply not acceptable.

libev

libev provides various different event notifications, to be able to get around edge triggered notifications, and still provide fair round-robin for all events we are going to have to build on top of multiple different events.

The events used will be:

• ev::io: This one is pretty self explanatory, this is for getting notified about input output changes. This is the one we are going to use on the ZMQ_FD.
• ev::prepare and ev::check: These two are generally used together, they can be used to change the event loop and or make modifications on the fly to events that have been registered with the event loop.
• ev::idle: This is an event that gets fired whenever the event loop has nothing else to do, so no other events fired, this will fire.

Plan of attack

Since the prepare and check run before and after the loop, we are going to be using those to do most of the work. We use an io so that we can turn off the idle when we can actually wait for a result from ZeroMQ's file descriptor, otherwise we use idle so that we will always get called once every loop.

In the prepare watcher callback we do the following:

1. Check to see what events ZeroMQ has for us, and check what events the user has requested.
2. If the ZeroMQ has an event for us that we want, and the user has requested that event, we start the idle watcher.
3. If ZeroMQ has no events, we start the io watcher.

In the check watcher callback we do the following:

1. Stop both the io and idle watchers, they were only there to make sure that our check watcher was called.
2. See what event ZeroMQ has for us, and check that against what the user wants. Depending on the event, call user defined function write() or user defined function read().
3. If this was a spurious wake-up on the part of ZeroMQ we simply ignore it and let libev go on to other events.

We could make all of this work by simply using the prepare, check and idle watchers, but that would mean libev would be busy-waiting for something to happen on the ZeroMQ socket. The io watcher is required simply so in times of nothing happening libev in its library can call into the kernels event handling mechanism and go to sleep. We can't use just the io watcher due to the edge-triggered notification, because we'd miss all kinds of ZeroMQ messages. So all four watchers are required, and play crucial parts in making this work.

Let's get down to code

Below you will find example code, it is not complete. Do note that I am using some C++11isms, error checking code may not be complete/correct and in general I don't suggest you copy and paste this without reading and understanding what it does.

The zmq_event class is meant to be used as a base class, inherit from it, and create the write() and read() functions. These functions will be called when you are able to read from the ZeroMQ socket, or when you are able to write to the ZeroMQ socket. You are guaranteed to be able to read one whole ZeroMQ message, so if it is a multi-part message, do make sure to loop on ZMQ_SNDMORE as required.

Upon instantiation it will automatically start being notified about events, we start off with ev::READ. When your sub-class wants to write to ZeroMQ it should put the messages to be written into a list somewhere, and set ev::READ | ev::WRITE on watcher_io, by calling watcher_io.set(socket_fd, ev::READ | ev::WRITE). write() will then be called, write a single message to ZeroMQ, and if necessary when finished writing, unset ev::WRITE using watcher_io.set(socket_fd, ev::READ). If you are not finished writing, after writing that singular message you may return and write() will be called again the next loop iteration. This way if you have a lot of data to write you don't starve the other events from receiving their notifications.

zmq_event.h

#include <string>

#include <ev++.h>
#include <zmq.hpp>

class zmq_event {
    public:
        zmq_event(zmq::context_t& context, int type, const std::string& connect);
        virtual ~zmq_event();

    protected:
        // This gets fired before the event loop, to prepare
        void before(ev::prepare& prep, int revents);

        // This is fired after the event loop, but before any other type of events
        void after(ev::check& check, int revents);

        // We need to have a no-op function available for those events that we
        // want to add to the list, but should never fire an actual event
        template <typename T>
            inline void noop(T& w, int revents) {};

        // Function we are going to call to write to the ZeroMQ socket
        virtual void write() = 0;

        // Function we are going to call to read from the ZeroMQ socket
        virtual void read() = 0;

        // Some helper function, one to start notifications
        void start_notify();

        // And one to stop notifications.
        void stop_notify();

        // Our event types
        ev::io      watcher_io;
        ev::prepare watcher_prepare;
        ev::check   watcher_check;
        ev::idle    watcher_idle;

        // Our ZeroMQ socket
        zmq::socket_t socket;
        int           socket_fd = -1;
};

zmq_event.cc

#include <stdexcept>

#include "zmq_event.h"

zmq_event::zmq_event(zmq::context_t& context, int type, const std::string& connect) : socket(context, type) {
    // Get the file descriptor for the socket
    size_t fd_len = sizeof(_socket_fd);
    socket.getsockopt(ZMQ_FD, &socket_fd, &fd_len);

    // Actually connect to the ZeroMQ endpoint, could replace this with a bind as well ...
    socket.bind(connect.c_str());

    // Set up all of our watchers

    // Have our IO watcher check for READ on the ZeroMQ socket
    watcher_io.set(socket_fd, ev::READ);

    // This watcher has a no-op callback
    watcher_io.set<zmq_event, &zmq_event::noop>(this);

    // Set up our prepare watcher to call the before() function
    watcher_prepare.set<zmq_event, &zmq_event::before>(this);

    // Set up the check watcher to call the after() function
    watcher_check.set<zmq_event, &zmq_event::after>(this);

    // Set up our idle watcher, once again a no-op
    watcher_idle.set<zmq_event, &zmq_event::noop>(this);

    // Tell libev to start notifying us!
    start_notify();
}

zmq_event::~zmq_event() {}

zmq_event::before(ev::prepare&, int revents) {
    if (EV_ERROR & revents) {
        throw std::runtime_error("libev error");
    }

    // Get any events that may be waiting
    uint32_t zevents = 0;
    size_t zevents_len = sizeof(zevents);

    // Lucky for us, getting the events available doesn't invalidate the
    // events, so that calling this in `before()` and in `after()` will
    // give us the same results.
    socket.getsockopt(ZMQ_EVENTS, &zevents, &zevents_len);

    // Check what events exists, and check it against what event we want. We
    // "abuse" our watcher_io.events for this information.
    if ((zevents & ZMQ_POLLOUT) && (watcher_io.events & ev::WRITE)) {
        watcher_idle.start();
        return;
    }

    if ((zevents & ZMQ_POLLIN) && (watcher_io.events & ev::READ)) {
        watcher_idle.start();
        return;
    }

    // No events ready to be processed, we'll just go watch some io
    watcher_io.start();
}

zmq_event::after(ev::check&, int revents) {
    if (EV_ERROR & revents) {
        throw std::runtime_error("libev error");
    }

    // Stop both the idle and the io watcher, no point in calling the no-op callback
    // One of them will be reactived by before() on the next loop
    watcher_idle.stop();
    watcher_io.stop();

    // Get the events
    uint32_t zevents = 0;
    size_t zevents_len = sizeof(zevents);
    socket.getsockopt(ZMQ_EVENTS, &zevents, &zevents_len);

    // Check the events and call the users read/write function
    if ((zevents & ZMQ_POLLIN) && (watcher_io.events & ev::READ)) {
        this->read();
    }

    if ((zevents & ZMQ_POLLOUT) && (watcher_io.events & ev::WRITE)) {
        this->write();
    }
}

zmq_event::start_notify() {
    watcher_check.start();
    watcher_prepare.start();
}

zmq_event::stop_notify() {
    watcher_check.stop();
    watcher_prepare.stop();
}

Other event loops

libev is but just one of many event loops that exist out there, hopefully this shows how it is possible to embed ZeroMQ into an event loop, thereby making it easier to embed ZeroMQ into any other event loops.

ZeroMQ - Edge Triggered Notification

ZeroMQ is an absolutely fantastic communications library that has allowed me to very easily create networks/connections between multiple different applications to communicate in an extremely fast yet directed manner and allow scaling of the connected entities with almost no second thought.

I have used ZeroMQ within Python and C++, and it has been fantastic leaving that part of the communication layer to a library rather than writing something similar myself. However, embedding ZeroMQ into an another event loop has proven to have some quirks. They are documented in the API, but not so much by other users/examples found around the web, and caught me by surprise.

This post mainly goes over the issue, the difference between edge and level triggered and how ZeroMQ works. I am hoping to spend some time documenting how to embed ZeroMQ correctly inside libev at a later date.

The quirk

ZeroMQ allows you to get an underlying file descriptor (ZMQ_FD) for a ZeroMQ socket using getsockopt(), this is not the actual ZeroMQ file descriptor but a stand-in:

int fd = 0;
size_t fd_len = sizeof(fd);
zmq_socket.getsockopt(ZMQ_FD, &fd, &fd_len);

This file descriptor can then be passed on to poll(), select(), kqueue() or whatever your event notification library of choice may be, the idea being that if the file descriptor is notified to be available for reading then you check the sockets ZMQ_EVENTS and find out if you are allowed to read or write to the ZeroMQ socket, or even possibly neither (false positive). The same file descriptor is also used internally within the ZeroMQ library for notification and thus we may receive an event notification that we can't do anything with.

As stated in the ZeroMQ API guide for getsockopt() under ZMQ_FD:

the ØMQ library shall signal any pending events on the socket in an edge-triggered fashion by making the file descriptor become ready for reading.

and further it states:

The ability to read from the returned file descriptor does not necessarily indicate that messages are available to be read from, or can be written to, the underlying socket; applications must retrieve the actual event state with a subsequent retrieval of the ZMQ_EVENTS option.

So far so good, so after we return from our event notification on the file descriptor we simply ask ZeroMQ whether or not we can read or write (or possibly both):

int zevents = 0;
size_t zevents_len = sizeof(zevents);
zmq_socket.getsockopt(ZMQ_EVENTS, &zevents, &zevents_len);

if (zevents & ZMQ_POLLIN) {
    // We can read from the ZeroMQ socket
}

if (zevents & ZMQ_POLLOUT) {
    // We can write to the ZeroMQ socket
}

// If neither of the above is true, then it was a false positive

So now you are happily reading from ZeroMQ and using the data as you wish, except you start to notice that sometimes it can take a little while for data to show up, or that if the end-point you are connected to on ZeroMQ sends a lot of data it is not delivered in a timely fashion. This is where the documentation for edge triggered comes into play, and that is something that is different about ZeroMQ compared to for example BSD sockets.

Edge triggered versus level triggered

Edge and level triggered refer to what kind of signal a device will output in an attempt to gain the attention of another device. Edge and level triggered have their history in electronics and hardware, where they play a very crucial role.

For both level and edge triggered assume the data is represented by this simple string:

00000000000111111111111111111111100000000011111111111000000000

The zeroes (0) are where no data is available, and the ones (1) are where data is available for processing.

Level triggered

Level triggered devices when they require attention, and until they no longer require attention will signal with a high voltage (going back to early electronics), and when they no longer require attention signal with a low voltage.

Data:  00000000001111111111111111111111000000000011111111111000000000
Block:      A               B              C         D
+ 5v              ____________________           _________
  -    __________|                    |_________|         |__________
  0v
State: 0          1                    0          1        0

At this point, once the signal turns to 1 in block B we can start processing data, and until we are done processing data the signal will stay 1. Once we are done processing the device will signal that we are done by setting the state to 0 in block C, and the process can repeat itself with block D.

Edge triggered

Edge triggered devices when they require attention will simply turn on the signal for a short period of time and then turn it back off until all of the requireding processing is complete at which point they will once again signal that they need processing.

Data:  00000000001111111111111111111111100000000011111111111000000000
Block:      A     B              C                 D
+ 5v              __                              __
  -    __________|  |____________________________|  |________________
  0v
State: 0          1              0                1        0

In this case when the trigger arrives you have to process all of the data available. When block B arrives and we get 1 as our signal, we have to process all of the data, we won't know when we are done other than some other notification scheme by the device. (Such as a special message that says we are done). Then when we are done we can go back to listening to waiting for a trigger.

If however in the time that we are processing the data from the trigger at B receive more data to process we won't get notified because we weren't officially done processing data from the device, and thus the device will not notify us of the new data again until we have processed all of the data.

There are some devices that as soon as you start processing their data, you re-enable the trigger, so if you receive new data while still processing the old you will get a new notification.

ZeroMQ uses edge triggered

When multiple message arrive for the ZeroMQ socket, one notification is sent. You check to see if the ZMQ_EVENTS actually contains ZMQ_POLLIN and then read a single message from ZeroMQ. Content that you have processed the data ZeroMQ has for you, you add the file descriptor back to the event notification library and wait. However, you know you sent two messages, but all you got was one. The next time you send data to the socket, it returns the second message, and you have now send three but the third doesn't seem to arrive.

This is where the edge triggered comes in. ZeroMQ only notifies you once, and after it has notified you it won't notify you again until it receives new messages, EVEN if you have messages waiting for you. In BSD sockets when you recv() data from the socket, if it is not all of the data it will simply tell you, and until all of the data has been read that the socket has to offer it is going to tell you (level triggered).

This means that when you get a notification from ZeroMQ you have to use a loop to process all of the messages that are coming your way:

int zevents = 0;
size_t zevents_len = sizeof(zevents);
zmq_socket.getsockopt(ZMQ_EVENTS, &zevents, &zevents_len);

do {
    if (zevents & ZMQ_POLLIN) {
        // We can read from the ZeroMQ socket
    } else {
        break;
    }

    // Check to see if there is more to read ...
    zmq_socket.getsockopt(ZMQ_EVENTS, &zevents, &zevents_len);
} while (zevents & ZMQ_POLLIN);

if (zevents & ZMQ_POLLOUT) {
    // We can write to the ZeroMQ socket
}

// If neither of the above is true, then it was a false positive

The same is true if you need to send multiple messages, you will need to check ZMQ_EVENTS for ZMQ_POLLOUT and send messages so long as that holds true, as soon as it is no longer true you will need to wait until you get triggered again by using the file descriptor in your event notification library.

Embedding or co-existing with another event loop

Embedding ZeroMQ into a different event loop becomes more difficult when you realise that it is edge triggered versus level triggered. If you want to allow fair processing of data between events it becomes harder to accomplish with ZeroMQ because you are required to process all of the data before going back to your event loop or requires the setting up of different types of events depending on the state that the ZeroMQ socket is in to make sure you process all of the data and don't forget any.

Edge triggered for the ZeroMQ file descriptor was a design decision which unfortunately made the use of ZeroMQ sockets more difficult and different from BSD sockets which makes it harder for application programmers to do the right thing. There are plenty of questions on the internet regarding the proper usage of ZMQ_FD with various different event notification schemes/implementations.