EXAMPLE PROGRAM

   // a single header file is required
   #include <ev.h>

   #include <stdio.h> // for puts

   // every watcher type has its own typedef'd struct
   // with the name ev_TYPE
   ev_io stdin_watcher;
   ev_timer timeout_watcher;

   // all watcher callbacks have a similar signature
   // this callback is called when data is readable on stdin
   static void
   stdin_cb (EV_P_ ev_io *w, int revents)
   {
     puts ("stdin ready");
     // for one-shot events, one must manually stop the watcher
     // with its corresponding stop function.
     ev_io_stop (EV_A_ w);

     // this causes all nested ev_run's to stop iterating
     ev_break (EV_A_ EVBREAK_ALL);
   }

   // another callback, this time for a time-out
   static void
   timeout_cb (EV_P_ ev_timer *w, int revents)
   {
     puts ("timeout");
     // this causes the innermost ev_run to stop iterating
     ev_break (EV_A_ EVBREAK_ONE);
   }

   int
   main (void)
   {
     // use the default event loop unless you have special needs
     struct ev_loop *loop = EV_DEFAULT;

     // initialise an io watcher, then start it
     // this one will watch for stdin to become readable
     ev_io_init (&stdin_watcher, stdin_cb, /*STDIN_FILENO*/ 0, EV_READ);
     ev_io_start (loop, &stdin_watcher);

     // initialise a timer watcher, then start it
     // simple non-repeating 5.5 second timeout
     ev_timer_init (&timeout_watcher, timeout_cb, 5.5, 0.);
     ev_timer_start (loop, &timeout_watcher);

     // now wait for events to arrive
     ev_run (loop, 0);

     // break was called, so exit
     return 0;
   }

FEATURES

It also is quite fast (see this benchmark <http://libev.schmorp.de/bench.html> comparing it to libevent for example).

CONVENTIONS

TIME REPRESENTATION

Unlike the name component "stamp" might indicate, it is also used for time differences (e.g. delays) throughout libev.

GENERIC WATCHER FUNCTIONS

"ev_init" (ev_TYPE *watcher, callback)

This macro initialises the generic portion of a watcher. The contents of the watcher object can be arbitrary (so "malloc" will do). Only the generic parts of the watcher are initialised, you need to call the type-specific "ev_TYPE_set" macro afterwards to initialise the type-specific parts. For each type there is also a "ev_TYPE_init" macro which rolls both calls into one.

You can reinitialise a watcher at any time as long as it has been stopped (or never started) and there are no pending events outstanding.

The callback is always of type "void (*)(struct ev_loop *loop, ev_TYPE *watcher, int revents)".

Example: Initialise an "ev_io" watcher in two steps.

   ev_io w;
   ev_init (&w, my_cb);
   ev_io_set (&w, STDIN_FILENO, EV_READ);
    

"ev_TYPE_set" (ev_TYPE *watcher, [args])

This macro initialises the type-specific parts of a watcher. You need to call "ev_init" at least once before you call this macro, but you can call "ev_TYPE_set" any number of times. You must not, however, call this macro on a watcher that is active (it can be pending, however, which is a difference to the "ev_init" macro).

Although some watcher types do not have type-specific arguments (e.g. "ev_prepare") you still need to call its "set" macro.

See "ev_init", above, for an example.

"ev_TYPE_init" (ev_TYPE *watcher, callback, [args])

This convenience macro rolls both "ev_init" and "ev_TYPE_set" macro calls into a single call. This is the most convenient method to initialise a watcher. The same limitations apply, of course.

Example: Initialise and set an "ev_io" watcher in one step.

   ev_io_init (&w, my_cb, STDIN_FILENO, EV_READ);
    

"ev_TYPE_start" (loop, ev_TYPE *watcher)

Starts (activates) the given watcher. Only active watchers will receive events. If the watcher is already active nothing will happen.

Example: Start the "ev_io" watcher that is being abused as example in this whole section.

   ev_io_start (EV_DEFAULT_UC, &w);
    

"ev_TYPE_stop" (loop, ev_TYPE *watcher)

Stops the given watcher if active, and clears the pending status (whether the watcher was active or not).

It is possible that stopped watchers are pending - for example, non-repeating timers are being stopped when they become pending - but calling "ev_TYPE_stop" ensures that the watcher is neither active nor pending. If you want to free or reuse the memory used by the watcher it is therefore a good idea to always call its "ev_TYPE_stop" function.

bool ev_is_active (ev_TYPE *watcher)

Returns a true value iff the watcher is active (i.e. it has been started and not yet been stopped). As long as a watcher is active you must not modify it.

bool ev_is_pending (ev_TYPE *watcher)

Returns a true value iff the watcher is pending, (i.e. it has outstanding events but its callback has not yet been invoked). As long as a watcher is pending (but not active) you must not call an init function on it (but "ev_TYPE_set" is safe), you must not change its priority, and you must make sure the watcher is available to libev (e.g. you cannot "free ()" it).

callback ev_cb (ev_TYPE *watcher)

Returns the callback currently set on the watcher.

ev_set_cb (ev_TYPE *watcher, callback)

Change the callback. You can change the callback at virtually any time (modulo threads).

ev_set_priority (ev_TYPE *watcher, int priority)

int ev_priority (ev_TYPE *watcher)

Set and query the priority of the watcher. The priority is a small integer between "EV_MAXPRI" (default: 2) and "EV_MINPRI" (default: "-2"). Pending watchers with higher priority will be invoked before watchers with lower priority, but priority will not keep watchers from being executed (except for "ev_idle" watchers).

If you need to suppress invocation when higher priority events are pending you need to look at "ev_idle" watchers, which provide this functionality.

You must not change the priority of a watcher as long as it is active or pending.

Setting a priority outside the range of "EV_MINPRI" to "EV_MAXPRI" is fine, as long as you do not mind that the priority value you query might or might not have been clamped to the valid range.

The default priority used by watchers when no priority has been set is always 0, which is supposed to not be too high and not be too low :).

See "WATCHER PRIORITY MODELS", below, for a more thorough treatment of priorities.

ev_invoke (loop, ev_TYPE *watcher, int revents)

Invoke the "watcher" with the given "loop" and "revents". Neither "loop" nor "revents" need to be valid as long as the watcher callback can deal with that fact, as both are simply passed through to the callback.

int ev_clear_pending (loop, ev_TYPE *watcher)

If the watcher is pending, this function clears its pending status and returns its "revents" bitset (as if its callback was invoked). If the watcher isn't pending it does nothing and returns 0.

Sometimes it can be useful to "poll" a watcher instead of waiting for its callback to be invoked, which can be accomplished with this function.

ev_feed_event (loop, ev_TYPE *watcher, int revents)

Feeds the given event set into the event loop, as if the specified event had happened for the specified watcher (which must be a pointer to an initialised but not necessarily started event watcher). Obviously you must not free the watcher as long as it has pending events.

Stopping the watcher, letting libev invoke it, or calling "ev_clear_pending" will clear the pending event, even if the watcher was not started in the first place.

See also "ev_feed_fd_event" and "ev_feed_signal_event" for related functions that do not need a watcher.

See also the "ASSOCIATING CUSTOM DATA WITH A WATCHER" and "BUILDING YOUR OWN COMPOSITE WATCHERS" idioms.

WATCHER STATES

initialised

Before a watcher can be registered with the event loop it has to be initialised. This can be done with a call to "ev_TYPE_init", or calls to "ev_init" followed by the watcher-specific "ev_TYPE_set" function.

In this state it is simply some block of memory that is suitable for use in an event loop. It can be moved around, freed, reused etc. at will - as long as you either keep the memory contents intact, or call "ev_TYPE_init" again.

started/running/active

Once a watcher has been started with a call to "ev_TYPE_start" it becomes property of the event loop, and is actively waiting for events. While in this state it cannot be accessed (except in a few documented ways), moved, freed or anything else - the only legal thing is to keep a pointer to it, and call libev functions on it that are documented to work on active watchers.

pending

If a watcher is active and libev determines that an event it is interested in has occurred (such as a timer expiring), it will become pending. It will stay in this pending state until either it is stopped or its callback is about to be invoked, so it is not normally pending inside the watcher callback.

The watcher might or might not be active while it is pending (for example, an expired non-repeating timer can be pending but no longer active). If it is stopped, it can be freely accessed (e.g. by calling "ev_TYPE_set"), but it is still property of the event loop at this time, so cannot be moved, freed or reused. And if it is active the rules described in the previous item still apply.

It is also possible to feed an event on a watcher that is not active (e.g. via "ev_feed_event"), in which case it becomes pending without being active.

stopped

A watcher can be stopped implicitly by libev (in which case it might still be pending), or explicitly by calling its "ev_TYPE_stop" function. The latter will clear any pending state the watcher might be in, regardless of whether it was active or not, so stopping a watcher explicitly before freeing it is often a good idea.

While stopped (and not pending) the watcher is essentially in the initialised state, that is, it can be reused, moved, modified in any way you wish (but when you trash the memory block, you need to "ev_TYPE_init" it again).

WATCHER PRIORITY MODELS

In libev, watcher priorities can be set using "ev_set_priority". See its description for the more technical details such as the actual priority range.

There are two common ways how these these priorities are being interpreted by event loops:

In the more common lock-out model, higher priorities "lock out" invocation of lower priority watchers, which means as long as higher priority watchers receive events, lower priority watchers are not being invoked.

The less common only-for-ordering model uses priorities solely to order callback invocation within a single event loop iteration: Higher priority watchers are invoked before lower priority ones, but they all get invoked before polling for new events.

Libev uses the second (only-for-ordering) model for all its watchers except for idle watchers (which use the lock-out model).

The rationale behind this is that implementing the lock-out model for watchers is not well supported by most kernel interfaces, and most event libraries will just poll for the same events again and again as long as their callbacks have not been executed, which is very inefficient in the common case of one high-priority watcher locking out a mass of lower priority ones.

Static (ordering) priorities are most useful when you have two or more watchers handling the same resource: a typical usage example is having an "ev_io" watcher to receive data, and an associated "ev_timer" to handle timeouts. Under load, data might be received while the program handles other jobs, but since timers normally get invoked first, the timeout handler will be executed before checking for data. In that case, giving the timer a lower priority than the I/O watcher ensures that I/O will be handled first even under adverse conditions (which is usually, but not always, what you want).

Since idle watchers use the "lock-out" model, meaning that idle watchers will only be executed when no same or higher priority watchers have received events, they can be used to implement the "lock-out" model when required.

For example, to emulate how many other event libraries handle priorities, you can associate an "ev_idle" watcher to each such watcher, and in the normal watcher callback, you just start the idle watcher. The real processing is done in the idle watcher callback. This causes libev to continuously poll and process kernel event data for the watcher, but when the lock-out case is known to be rare (which in turn is rare :), this is workable.

Usually, however, the lock-out model implemented that way will perform miserably under the type of load it was designed to handle. In that case, it might be preferable to stop the real watcher before starting the idle watcher, so the kernel will not have to process the event in case the actual processing will be delayed for considerable time.

Here is an example of an I/O watcher that should run at a strictly lower priority than the default, and which should only process data when no other events are pending:

   ev_idle idle; // actual processing watcher
   ev_io io;     // actual event watcher

   static void
   io_cb (EV_P_ ev_io *w, int revents)
   {
     // stop the I/O watcher, we received the event, but
     // are not yet ready to handle it.
     ev_io_stop (EV_A_ w);

     // start the idle watcher to handle the actual event.
     // it will not be executed as long as other watchers
     // with the default priority are receiving events.
     ev_idle_start (EV_A_ &idle);
   }

   static void
   idle_cb (EV_P_ ev_idle *w, int revents)
   {
     // actual processing
     read (STDIN_FILENO, ...);

     // have to start the I/O watcher again, as
     // we have handled the event
     ev_io_start (EV_P_ &io);
   }

   // initialisation
   ev_idle_init (&idle, idle_cb);
   ev_io_init (&io, io_cb, STDIN_FILENO, EV_READ);
   ev_io_start (EV_DEFAULT_ &io);

In the "real" world, it might also be beneficial to start a timer, so that low-priority connections can not be locked out forever under load. This enables your program to keep a lower latency for important connections during short periods of high load, while not completely locking out less important ones.

"ev_io" - is this file descriptor readable or writable?

In general you can register as many read and/or write event watchers per fd as you want (as long as you don't confuse yourself). Setting all file descriptors to non-blocking mode is also usually a good idea (but not required if you know what you are doing).

Another thing you have to watch out for is that it is quite easy to receive "spurious" readiness notifications, that is, your callback might be called with "EV_READ" but a subsequent "read"(2) will actually block because there is no data. It is very easy to get into this situation even with a relatively standard program structure. Thus it is best to always use non-blocking I/O: An extra "read"(2) returning "EAGAIN" is far preferable to a program hanging until some data arrives.

If you cannot run the fd in non-blocking mode (for example you should not play around with an Xlib connection), then you have to separately re-test whether a file descriptor is really ready with a known-to-be good interface such as poll (fortunately in the case of Xlib, it already does this on its own, so its quite safe to use). Some people additionally use "SIGALRM" and an interval timer, just to be sure you won't block indefinitely.

But really, best use non-blocking mode.

The special problem of disappearing file descriptors

Some backends (e.g. kqueue, epoll, linuxaio) need to be told about closing a file descriptor (either due to calling "close" explicitly or any other means, such as "dup2"). The reason is that you register interest in some file descriptor, but when it goes away, the operating system will silently drop this interest. If another file descriptor with the same number then is registered with libev, there is no efficient way to see that this is, in fact, a different file descriptor.

To avoid having to explicitly tell libev about such cases, libev follows the following policy: Each time "ev_io_set" is being called, libev will assume that this is potentially a new file descriptor, otherwise it is assumed that the file descriptor stays the same. That means that you have to call "ev_io_set" (or "ev_io_init") when you change the descriptor even if the file descriptor number itself did not change.

This is how one would do it normally anyway, the important point is that the libev application should not optimise around libev but should leave optimisations to libev.

The special problem of dup'ed file descriptors

Some backends (e.g. epoll), cannot register events for file descriptors, but only events for the underlying file descriptions. That means when you have "dup ()"'ed file descriptors or weirder constellations, and register events for them, only one file descriptor might actually receive events.

There is no workaround possible except not registering events for potentially "dup ()"'ed file descriptors, or to resort to "EVBACKEND_SELECT" or "EVBACKEND_POLL".

The special problem of files

Many people try to use "select" (or libev) on file descriptors representing files, and expect it to become ready when their program doesn't block on disk accesses (which can take a long time on their own).

However, this cannot ever work in the "expected" way - you get a readiness notification as soon as the kernel knows whether and how much data is there, and in the case of open files, that's always the case, so you always get a readiness notification instantly, and your read (or possibly write) will still block on the disk I/O.

Another way to view it is that in the case of sockets, pipes, character devices and so on, there is another party (the sender) that delivers data on its own, but in the case of files, there is no such thing: the disk will not send data on its own, simply because it doesn't know what you wish to read - you would first have to request some data.

Since files are typically not-so-well supported by advanced notification mechanism, libev tries hard to emulate POSIX behaviour with respect to files, even though you should not use it. The reason for this is convenience: sometimes you want to watch STDIN or STDOUT, which is usually a tty, often a pipe, but also sometimes files or special devices (for example, "epoll" on Linux works with /dev/random but not with /dev/urandom), and even though the file might better be served with asynchronous I/O instead of with non-blocking I/O, it is still useful when it "just works" instead of freezing.

So avoid file descriptors pointing to files when you know it (e.g. use libeio), but use them when it is convenient, e.g. for STDIN/STDOUT, or when you rarely read from a file instead of from a socket, and want to reuse the same code path.

The special problem of fork

Some backends (epoll, kqueue, linuxaio, iouring) do not support "fork ()" at all or exhibit useless behaviour. Libev fully supports fork, but needs to be told about it in the child if you want to continue to use it in the child.

To support fork in your child processes, you have to call "ev_loop_fork ()" after a fork in the child, enable "EVFLAG_FORKCHECK", or resort to "EVBACKEND_SELECT" or "EVBACKEND_POLL".

The special problem of SIGPIPE

While not really specific to libev, it is easy to forget about "SIGPIPE": when writing to a pipe whose other end has been closed, your program gets sent a SIGPIPE, which, by default, aborts your program. For most programs this is sensible behaviour, for daemons, this is usually undesirable.

So when you encounter spurious, unexplained daemon exits, make sure you ignore SIGPIPE (and maybe make sure you log the exit status of your daemon somewhere, as that would have given you a big clue).

The special problem of accept()ing when you can't

Many implementations of the POSIX "accept" function (for example, found in post-2004 Linux) have the peculiar behaviour of not removing a connection from the pending queue in all error cases.

For example, larger servers often run out of file descriptors (because of resource limits), causing "accept" to fail with "ENFILE" but not rejecting the connection, leading to libev signalling readiness on the next iteration again (the connection still exists after all), and typically causing the program to loop at 100% CPU usage.

Unfortunately, the set of errors that cause this issue differs between operating systems, there is usually little the app can do to remedy the situation, and no known thread-safe method of removing the connection to cope with overload is known (to me).

One of the easiest ways to handle this situation is to just ignore it - when the program encounters an overload, it will just loop until the situation is over. While this is a form of busy waiting, no OS offers an event-based way to handle this situation, so it's the best one can do.

A better way to handle the situation is to log any errors other than "EAGAIN" and "EWOULDBLOCK", making sure not to flood the log with such messages, and continue as usual, which at least gives the user an idea of what could be wrong ("raise the ulimit!"). For extra points one could stop the "ev_io" watcher on the listening fd "for a while", which reduces CPU usage.

If your program is single-threaded, then you could also keep a dummy file descriptor for overload situations (e.g. by opening /dev/null), and when you run into "ENFILE" or "EMFILE", close it, run "accept", close that fd, and create a new dummy fd. This will gracefully refuse clients under typical overload conditions.

The last way to handle it is to simply log the error and "exit", as is often done with "malloc" failures, but this results in an easy opportunity for a DoS attack.

Watcher-Specific Functions

ev_io_init (ev_io *, callback, int fd, int events)

ev_io_set (ev_io *, int fd, int events)

Configures an "ev_io" watcher. The "fd" is the file descriptor to receive events for and "events" is either "EV_READ", "EV_WRITE", both "EV_READ | EV_WRITE" or 0, to express the desire to receive the given events.

Note that setting the "events" to 0 and starting the watcher is supported, but not specially optimized - if your program sometimes happens to generate this combination this is fine, but if it is easy to avoid starting an io watcher watching for no events you should do so.

ev_io_modify (ev_io *, int events)

Similar to "ev_io_set", but only changes the requested events. Using this might be faster with some backends, as libev can assume that the "fd" still refers to the same underlying file description, something it cannot do when using "ev_io_set".

int fd [no-modify]

The file descriptor being watched. While it can be read at any time, you must not modify this member even when the watcher is stopped - always use "ev_io_set" for that.

int events [no-modify]

The set of events the fd is being watched for, among other flags. Remember that this is a bit set - to test for "EV_READ", use "w->events & EV_READ", and similarly for "EV_WRITE".

As with "fd", you must not modify this member even when the watcher is stopped, always use "ev_io_set" or "ev_io_modify" for that.

Examples

Example: Call "stdin_readable_cb" when STDIN_FILENO has become, well readable, but only once. Since it is likely line-buffered, you could attempt to read a whole line in the callback.

   static void
   stdin_readable_cb (struct ev_loop *loop, ev_io *w, int revents)
   {
      ev_io_stop (loop, w);
     .. read from stdin here (or from w->fd) and handle any I/O errors
   }

   ...
   struct ev_loop *loop = ev_default_init (0);
   ev_io stdin_readable;
   ev_io_init (&stdin_readable, stdin_readable_cb, STDIN_FILENO, EV_READ);
   ev_io_start (loop, &stdin_readable);
   ev_run (loop, 0);

"ev_timer" - relative and optionally repeating timeouts

The timers are based on real time, that is, if you register an event that times out after an hour and you reset your system clock to January last year, it will still time out after (roughly) one hour. "Roughly" because detecting time jumps is hard, and some inaccuracies are unavoidable (the monotonic clock option helps a lot here).

The callback is guaranteed to be invoked only after its timeout has passed (not at, so on systems with very low-resolution clocks this might introduce a small delay, see "the special problem of being too early", below). If multiple timers become ready during the same loop iteration then the ones with earlier time-out values are invoked before ones of the same priority with later time-out values (but this is no longer true when a callback calls "ev_run" recursively).

Be smart about timeouts

Many real-world problems involve some kind of timeout, usually for error recovery. A typical example is an HTTP request - if the other side hangs, you want to raise some error after a while.

What follows are some ways to handle this problem, from obvious and inefficient to smart and efficient.

In the following, a 60 second activity timeout is assumed - a timeout that gets reset to 60 seconds each time there is activity (e.g. each time some data or other life sign was received).

1. Use a timer and stop, reinitialise and start it on activity.

This is the most obvious, but not the most simple way: In the beginning, start the watcher:

   ev_timer_init (timer, callback, 60., 0.);
   ev_timer_start (loop, timer);
    

Then, each time there is some activity, "ev_timer_stop" it, initialise it and start it again:

   ev_timer_stop (loop, timer);
   ev_timer_set (timer, 60., 0.);
   ev_timer_start (loop, timer);
    

This is relatively simple to implement, but means that each time there is some activity, libev will first have to remove the timer from its internal data structure and then add it again. Libev tries to be fast, but it's still not a constant-time operation.

2. Use a timer and re-start it with "ev_timer_again" inactivity.

This is the easiest way, and involves using "ev_timer_again" instead of "ev_timer_start".

To implement this, configure an "ev_timer" with a "repeat" value of 60 and then call "ev_timer_again" at start and each time you successfully read or write some data. If you go into an idle state where you do not expect data to travel on the socket, you can "ev_timer_stop" the timer, and "ev_timer_again" will automatically restart it if need be.

That means you can ignore both the "ev_timer_start" function and the "after" argument to "ev_timer_set", and only ever use the "repeat" member and "ev_timer_again".

At start:

   ev_init (timer, callback);
   timer->repeat = 60.;
   ev_timer_again (loop, timer);
    

Each time there is some activity:

   ev_timer_again (loop, timer);
    

It is even possible to change the time-out on the fly, regardless of whether the watcher is active or not:

   timer->repeat = 30.;
   ev_timer_again (loop, timer);
    

This is slightly more efficient then stopping/starting the timer each time you want to modify its timeout value, as libev does not have to completely remove and re-insert the timer from/into its internal data structure.

It is, however, even simpler than the "obvious" way to do it.

3. Let the timer time out, but then re-arm it as required.

This method is more tricky, but usually most efficient: Most timeouts are relatively long compared to the intervals between other activity - in our example, within 60 seconds, there are usually many I/O events with associated activity resets.

In this case, it would be more efficient to leave the "ev_timer" alone, but remember the time of last activity, and check for a real timeout only within the callback:

   ev_tstamp timeout = 60.;
   ev_tstamp last_activity; // time of last activity
   ev_timer timer;

   static void
   callback (EV_P_ ev_timer *w, int revents)
   {
     // calculate when the timeout would happen
     ev_tstamp after = last_activity - ev_now (EV_A) + timeout;

     // if negative, it means we the timeout already occurred
     if (after < 0.)
       {
         // timeout occurred, take action
       }
     else
       {
         // callback was invoked, but there was some recent 
         // activity. simply restart the timer to time out
         // after "after" seconds, which is the earliest time
         // the timeout can occur.
         ev_timer_set (w, after, 0.);
         ev_timer_start (EV_A_ w);
       }
   }
    

To summarise the callback: first calculate in how many seconds the timeout will occur (by calculating the absolute time when it would occur, "last_activity + timeout", and subtracting the current time, "ev_now (EV_A)" from that).

If this value is negative, then we are already past the timeout, i.e. we timed out, and need to do whatever is needed in this case.

Otherwise, we now the earliest time at which the timeout would trigger, and simply start the timer with this timeout value.

In other words, each time the callback is invoked it will check whether the timeout occurred. If not, it will simply reschedule itself to check again at the earliest time it could time out. Rinse. Repeat.

This scheme causes more callback invocations (about one every 60 seconds minus half the average time between activity), but virtually no calls to libev to change the timeout.

To start the machinery, simply initialise the watcher and set "last_activity" to the current time (meaning there was some activity just now), then call the callback, which will "do the right thing" and start the timer:

   last_activity = ev_now (EV_A);
   ev_init (&timer, callback);
   callback (EV_A_ &timer, 0);
    

When there is some activity, simply store the current time in "last_activity", no libev calls at all:

   if (activity detected)
     last_activity = ev_now (EV_A);
    

When your timeout value changes, then the timeout can be changed by simply providing a new value, stopping the timer and calling the callback, which will again do the right thing (for example, time out immediately :).

   timeout = new_value;
   ev_timer_stop (EV_A_ &timer);
   callback (EV_A_ &timer, 0);
    

This technique is slightly more complex, but in most cases where the time-out is unlikely to be triggered, much more efficient.

4. Wee, just use a double-linked list for your timeouts.

If there is not one request, but many thousands (millions...), all employing some kind of timeout with the same timeout value, then one can do even better:

When starting the timeout, calculate the timeout value and put the timeout at the end of the list.

Then use an "ev_timer" to fire when the timeout at the beginning of the list is expected to fire (for example, using the technique #3).

When there is some activity, remove the timer from the list, recalculate the timeout, append it to the end of the list again, and make sure to update the "ev_timer" if it was taken from the beginning of the list.

This way, one can manage an unlimited number of timeouts in O(1) time for starting, stopping and updating the timers, at the expense of a major complication, and having to use a constant timeout. The constant timeout ensures that the list stays sorted.

So which method the best?

Method #2 is a simple no-brain-required solution that is adequate in most situations. Method #3 requires a bit more thinking, but handles many cases better, and isn't very complicated either. In most case, choosing either one is fine, with #3 being better in typical situations.

Method #1 is almost always a bad idea, and buys you nothing. Method #4 is rather complicated, but extremely efficient, something that really pays off after the first million or so of active timers, i.e. it's usually overkill :)

The special problem of being too early

If you ask a timer to call your callback after three seconds, then you expect it to be invoked after three seconds - but of course, this cannot be guaranteed to infinite precision. Less obviously, it cannot be guaranteed to any precision by libev - imagine somebody suspending the process with a STOP signal for a few hours for example.

So, libev tries to invoke your callback as soon as possible after the delay has occurred, but cannot guarantee this.

A less obvious failure mode is calling your callback too early: many event loops compare timestamps with a "elapsed delay >= requested delay", but this can cause your callback to be invoked much earlier than you would expect.

To see why, imagine a system with a clock that only offers full second resolution (think windows if you can't come up with a broken enough OS yourself). If you schedule a one-second timer at the time 500.9, then the event loop will schedule your timeout to elapse at a system time of 500 (500.9 truncated to the resolution) + 1, or 501.

If an event library looks at the timeout 0.1s later, it will see "501 >= 501" and invoke the callback 0.1s after it was started, even though a one-second delay was requested - this is being "too early", despite best intentions.

This is the reason why libev will never invoke the callback if the elapsed delay equals the requested delay, but only when the elapsed delay is larger than the requested delay. In the example above, libev would only invoke the callback at system time 502, or 1.1s after the timer was started.

So, while libev cannot guarantee that your callback will be invoked exactly when requested, it can and does guarantee that the requested delay has actually elapsed, or in other words, it always errs on the "too late" side of things.

The special problem of time updates

Establishing the current time is a costly operation (it usually takes at least one system call): EV therefore updates its idea of the current time only before and after "ev_run" collects new events, which causes a growing difference between "ev_now ()" and "ev_time ()" when handling lots of events in one iteration.

The relative timeouts are calculated relative to the "ev_now ()" time. This is usually the right thing as this timestamp refers to the time of the event triggering whatever timeout you are modifying/starting. If you suspect event processing to be delayed and you need to base the timeout on the current time, use something like the following to adjust for it:

   ev_timer_set (&timer, after + (ev_time () - ev_now ()), 0.);

If the event loop is suspended for a long time, you can also force an update of the time returned by "ev_now ()" by calling "ev_now_update ()", although that will push the event time of all outstanding events further into the future.

The special problem of unsynchronised clocks

Modern systems have a variety of clocks - libev itself uses the normal "wall clock" clock and, if available, the monotonic clock (to avoid time jumps).

Neither of these clocks is synchronised with each other or any other clock on the system, so "ev_time ()" might return a considerably different time than "gettimeofday ()" or "time ()". On a GNU/Linux system, for example, a call to "gettimeofday" might return a second count that is one higher than a directly following call to "time".

The moral of this is to only compare libev-related timestamps with "ev_time ()" and "ev_now ()", at least if you want better precision than a second or so.

One more problem arises due to this lack of synchronisation: if libev uses the system monotonic clock and you compare timestamps from "ev_time" or "ev_now" from when you started your timer and when your callback is invoked, you will find that sometimes the callback is a bit "early".

This is because "ev_timer"s work in real time, not wall clock time, so libev makes sure your callback is not invoked before the delay happened, measured according to the real time, not the system clock.

If your timeouts are based on a physical timescale (e.g. "time out this connection after 100 seconds") then this shouldn't bother you as it is exactly the right behaviour.

If you want to compare wall clock/system timestamps to your timers, then you need to use "ev_periodic"s, as these are based on the wall clock time, where your comparisons will always generate correct results.

The special problems of suspended animation

When you leave the server world it is quite customary to hit machines that can suspend/hibernate - what happens to the clocks during such a suspend?

Some quick tests made with a Linux 2.6.28 indicate that a suspend freezes all processes, while the clocks ("times", "CLOCK_MONOTONIC") continue to run until the system is suspended, but they will not advance while the system is suspended. That means, on resume, it will be as if the program was frozen for a few seconds, but the suspend time will not be counted towards "ev_timer" when a monotonic clock source is used. The real time clock advanced as expected, but if it is used as sole clocksource, then a long suspend would be detected as a time jump by libev, and timers would be adjusted accordingly.

I would not be surprised to see different behaviour in different between operating systems, OS versions or even different hardware.

The other form of suspend (job control, or sending a SIGSTOP) will see a time jump in the monotonic clocks and the realtime clock. If the program is suspended for a very long time, and monotonic clock sources are in use, then you can expect "ev_timer"s to expire as the full suspension time will be counted towards the timers. When no monotonic clock source is in use, then libev will again assume a timejump and adjust accordingly.

It might be beneficial for this latter case to call "ev_suspend" and "ev_resume" in code that handles "SIGTSTP", to at least get deterministic behaviour in this case (you can do nothing against "SIGSTOP").

Watcher-Specific Functions and Data Members

ev_timer_init (ev_timer *, callback, ev_tstamp after, ev_tstamp repeat)

ev_timer_set (ev_timer *, ev_tstamp after, ev_tstamp repeat)

Configure the timer to trigger after "after" seconds (fractional and negative values are supported). If "repeat" is 0., then it will automatically be stopped once the timeout is reached. If it is positive, then the timer will automatically be configured to trigger again "repeat" seconds later, again, and again, until stopped manually.

The timer itself will do a best-effort at avoiding drift, that is, if you configure a timer to trigger every 10 seconds, then it will normally trigger at exactly 10 second intervals. If, however, your program cannot keep up with the timer (because it takes longer than those 10 seconds to do stuff) the timer will not fire more than once per event loop iteration.

ev_timer_again (loop, ev_timer *)

This will act as if the timer timed out, and restarts it again if it is repeating. It basically works like calling "ev_timer_stop", updating the timeout to the "repeat" value and calling "ev_timer_start".

The exact semantics are as in the following rules, all of which will be applied to the watcher:

ev_tstamp ev_timer_remaining (loop, ev_timer *)

Returns the remaining time until a timer fires. If the timer is active, then this time is relative to the current event loop time, otherwise it's the timeout value currently configured.

That is, after an "ev_timer_set (w, 5, 7)", "ev_timer_remaining" returns 5. When the timer is started and one second passes, "ev_timer_remaining" will return 4. When the timer expires and is restarted, it will return roughly 7 (likely slightly less as callback invocation takes some time, too), and so on.

ev_tstamp repeat [read-write]

The current "repeat" value. Will be used each time the watcher times out or "ev_timer_again" is called, and determines the next timeout (if any), which is also when any modifications are taken into account.

Examples

Example: Create a timer that fires after 60 seconds.

   static void
   one_minute_cb (struct ev_loop *loop, ev_timer *w, int revents)
   {
     .. one minute over, w is actually stopped right here
   }

   ev_timer mytimer;
   ev_timer_init (&mytimer, one_minute_cb, 60., 0.);
   ev_timer_start (loop, &mytimer);

Example: Create a timeout timer that times out after 10 seconds of inactivity.

   static void
   timeout_cb (struct ev_loop *loop, ev_timer *w, int revents)
   {
     .. ten seconds without any activity
   }

   ev_timer mytimer;
   ev_timer_init (&mytimer, timeout_cb, 0., 10.); /* note, only repeat used */
   ev_timer_again (&mytimer); /* start timer */
   ev_run (loop, 0);

   // and in some piece of code that gets executed on any "activity":
   // reset the timeout to start ticking again at 10 seconds
   ev_timer_again (&mytimer);

"ev_periodic" - to cron or not to cron?

Unlike "ev_timer", periodic watchers are not based on real time (or relative time, the physical time that passes) but on wall clock time (absolute time, the thing you can read on your calendar or clock). The difference is that wall clock time can run faster or slower than real time, and time jumps are not uncommon (e.g. when you adjust your wrist-watch).

You can tell a periodic watcher to trigger after some specific point in time: for example, if you tell a periodic watcher to trigger "in 10 seconds" (by specifying e.g. "ev_now () + 10.", that is, an absolute time not a delay) and then reset your system clock to January of the previous year, then it will take a year or more to trigger the event (unlike an "ev_timer", which would still trigger roughly 10 seconds after starting it, as it uses a relative timeout).

"ev_periodic" watchers can also be used to implement vastly more complex timers, such as triggering an event on each "midnight, local time", or other complicated rules. This cannot easily be done with "ev_timer" watchers, as those cannot react to time jumps.

As with timers, the callback is guaranteed to be invoked only when the point in time where it is supposed to trigger has passed. If multiple timers become ready during the same loop iteration then the ones with earlier time-out values are invoked before ones with later time-out values (but this is no longer true when a callback calls "ev_run" recursively).

Watcher-Specific Functions and Data Members

ev_periodic_init (ev_periodic *, callback, ev_tstamp offset, ev_tstamp interval, reschedule_cb)

ev_periodic_set (ev_periodic *, ev_tstamp offset, ev_tstamp interval, reschedule_cb)

Lots of arguments, let's sort it out... There are basically three modes of operation, and we will explain them from simplest to most complex:

ev_periodic_again (loop, ev_periodic *)

Simply stops and restarts the periodic watcher again. This is only useful when you changed some parameters or the reschedule callback would return a different time than the last time it was called (e.g. in a crond like program when the crontabs have changed).

ev_tstamp ev_periodic_at (ev_periodic *)

When active, returns the absolute time that the watcher is supposed to trigger next. This is not the same as the "offset" argument to "ev_periodic_set", but indeed works even in interval and manual rescheduling modes.

ev_tstamp offset [read-write]

When repeating, this contains the offset value, otherwise this is the absolute point in time (the "offset" value passed to "ev_periodic_set", although libev might modify this value for better numerical stability).

Can be modified any time, but changes only take effect when the periodic timer fires or "ev_periodic_again" is being called.

ev_tstamp interval [read-write]

The current interval value. Can be modified any time, but changes only take effect when the periodic timer fires or "ev_periodic_again" is being called.

ev_tstamp (*reschedule_cb)(ev_periodic *w, ev_tstamp now) [read-write]

The current reschedule callback, or 0, if this functionality is switched off. Can be changed any time, but changes only take effect when the periodic timer fires or "ev_periodic_again" is being called.

Examples

Example: Call a callback every hour, or, more precisely, whenever the system time is divisible by 3600. The callback invocation times have potentially a lot of jitter, but good long-term stability.

   static void
   clock_cb (struct ev_loop *loop, ev_periodic *w, int revents)
   {
     ... its now a full hour (UTC, or TAI or whatever your clock follows)
   }

   ev_periodic hourly_tick;
   ev_periodic_init (&hourly_tick, clock_cb, 0., 3600., 0);
   ev_periodic_start (loop, &hourly_tick);

Example: The same as above, but use a reschedule callback to do it:

   #include <math.h>

   static ev_tstamp
   my_scheduler_cb (ev_periodic *w, ev_tstamp now)
   {
     return now + (3600. - fmod (now, 3600.));
   }

   ev_periodic_init (&hourly_tick, clock_cb, 0., 0., my_scheduler_cb);

Example: Call a callback every hour, starting now:

   ev_periodic hourly_tick;
   ev_periodic_init (&hourly_tick, clock_cb,
                     fmod (ev_now (loop), 3600.), 3600., 0);
   ev_periodic_start (loop, &hourly_tick);

"ev_signal" - signal me when a signal gets signalled!

If you want signals to be delivered truly asynchronously, just use "sigaction" as you would do without libev and forget about sharing the signal. You can even use "ev_async" from a signal handler to synchronously wake up an event loop.

You can configure as many watchers as you like for the same signal, but only within the same loop, i.e. you can watch for "SIGINT" in your default loop and for "SIGIO" in another loop, but you cannot watch for "SIGINT" in both the default loop and another loop at the same time. At the moment, "SIGCHLD" is permanently tied to the default loop.

Only after the first watcher for a signal is started will libev actually register something with the kernel. It thus coexists with your own signal handlers as long as you don't register any with libev for the same signal.

If possible and supported, libev will install its handlers with "SA_RESTART" (or equivalent) behaviour enabled, so system calls should not be unduly interrupted. If you have a problem with system calls getting interrupted by signals you can block all signals in an "ev_check" watcher and unblock them in an "ev_prepare" watcher.

The special problem of inheritance over fork/execve/pthread_create

Both the signal mask ("sigprocmask") and the signal disposition ("sigaction") are unspecified after starting a signal watcher (and after stopping it again), that is, libev might or might not block the signal, and might or might not set or restore the installed signal handler (but see "EVFLAG_NOSIGMASK").

While this does not matter for the signal disposition (libev never sets signals to "SIG_IGN", so handlers will be reset to "SIG_DFL" on "execve"), this matters for the signal mask: many programs do not expect certain signals to be blocked.

This means that before calling "exec" (from the child) you should reset the signal mask to whatever "default" you expect (all clear is a good choice usually).

The simplest way to ensure that the signal mask is reset in the child is to install a fork handler with "pthread_atfork" that resets it. That will catch fork calls done by libraries (such as the libc) as well.

In current versions of libev, the signal will not be blocked indefinitely unless you use the "signalfd" API ("EV_SIGNALFD"). While this reduces the window of opportunity for problems, it will not go away, as libev has to modify the signal mask, at least temporarily.

So I can't stress this enough: If you do not reset your signal mask when you expect it to be empty, you have a race condition in your code. This is not a libev-specific thing, this is true for most event libraries.

The special problem of threads signal handling

POSIX threads has problematic signal handling semantics, specifically, a lot of functionality (sigfd, sigwait etc.) only really works if all threads in a process block signals, which is hard to achieve.

When you want to use sigwait (or mix libev signal handling with your own for the same signals), you can tackle this problem by globally blocking all signals before creating any threads (or creating them with a fully set sigprocmask) and also specifying the "EVFLAG_NOSIGMASK" when creating loops. Then designate one thread as "signal receiver thread" which handles these signals. You can pass on any signals that libev might be interested in by calling "ev_feed_signal".

Watcher-Specific Functions and Data Members

ev_signal_init (ev_signal *, callback, int signum)

ev_signal_set (ev_signal *, int signum)

Configures the watcher to trigger on the given signal number (usually one of the "SIGxxx" constants).

int signum [read-only]

The signal the watcher watches out for.

Examples

Example: Try to exit cleanly on SIGINT.

   static void
   sigint_cb (struct ev_loop *loop, ev_signal *w, int revents)
   {
     ev_break (loop, EVBREAK_ALL);
   }

   ev_signal signal_watcher;
   ev_signal_init (&signal_watcher, sigint_cb, SIGINT);
   ev_signal_start (loop, &signal_watcher);

"ev_child" - watch out for process status changes

Only the default event loop is capable of handling signals, and therefore you can only register child watchers in the default event loop.

Due to some design glitches inside libev, child watchers will always be handled at maximum priority (their priority is set to "EV_MAXPRI" by libev)

Process Interaction

Libev grabs "SIGCHLD" as soon as the default event loop is initialised. This is necessary to guarantee proper behaviour even if the first child watcher is started after the child exits. The occurrence of "SIGCHLD" is recorded asynchronously, but child reaping is done synchronously as part of the event loop processing. Libev always reaps all children, even ones not watched.

Overriding the Built-In Processing

Libev offers no special support for overriding the built-in child processing, but if your application collides with libev's default child handler, you can override it easily by installing your own handler for "SIGCHLD" after initialising the default loop, and making sure the default loop never gets destroyed. You are encouraged, however, to use an event-based approach to child reaping and thus use libev's support for that, so other libev users can use "ev_child" watchers freely.

Stopping the Child Watcher

Currently, the child watcher never gets stopped, even when the child terminates, so normally one needs to stop the watcher in the callback. Future versions of libev might stop the watcher automatically when a child exit is detected (calling "ev_child_stop" twice is not a problem).

Watcher-Specific Functions and Data Members

ev_child_init (ev_child *, callback, int pid, int trace)

ev_child_set (ev_child *, int pid, int trace)

Configures the watcher to wait for status changes of process "pid" (or any process if "pid" is specified as 0). The callback can look at the "rstatus" member of the "ev_child" watcher structure to see the status word (use the macros from "sys/wait.h" and see your systems "waitpid" documentation). The "rpid" member contains the pid of the process causing the status change. "trace" must be either 0 (only activate the watcher when the process terminates) or 1 (additionally activate the watcher when the process is stopped or continued).

int pid [read-only]

The process id this watcher watches out for, or 0, meaning any process id.

int rpid [read-write]

The process id that detected a status change.

int rstatus [read-write]

The process exit/trace status caused by "rpid" (see your systems "waitpid" and "sys/wait.h" documentation for details).

Examples

Example: "fork()" a new process and install a child handler to wait for its completion.

   ev_child cw;

   static void
   child_cb (EV_P_ ev_child *w, int revents)
   {
     ev_child_stop (EV_A_ w);
     printf ("process %d exited with status %x\n", w->rpid, w->rstatus);
   }

   pid_t pid = fork ();

   if (pid < 0)
     // error
   else if (pid == 0)
     {
       // the forked child executes here
       exit (1);
     }
   else
     {
       ev_child_init (&cw, child_cb, pid, 0);
       ev_child_start (EV_DEFAULT_ &cw);
     }

"ev_stat" - did the file attributes just change?

The path does not need to exist: changing from "path exists" to "path does not exist" is a status change like any other. The condition "path does not exist" (or more correctly "path cannot be stat'ed") is signified by the "st_nlink" field being zero (which is otherwise always forced to be at least one) and all the other fields of the stat buffer having unspecified contents.

The path must not end in a slash or contain special components such as "." or "..". The path should be absolute: If it is relative and your working directory changes, then the behaviour is undefined.

Since there is no portable change notification interface available, the portable implementation simply calls stat(2) regularly on the path to see if it changed somehow. You can specify a recommended polling interval for this case. If you specify a polling interval of 0 (highly recommended!) then a suitable, unspecified default value will be used (which you can expect to be around five seconds, although this might change dynamically). Libev will also impose a minimum interval which is currently around 0.1, but that's usually overkill.

This watcher type is not meant for massive numbers of stat watchers, as even with OS-supported change notifications, this can be resource-intensive.

At the time of this writing, the only OS-specific interface implemented is the Linux inotify interface (implementing kqueue support is left as an exercise for the reader. Note, however, that the author sees no way of implementing "ev_stat" semantics with kqueue, except as a hint).

ABI Issues (Largefile Support)

Libev by default (unless the user overrides this) uses the default compilation environment, which means that on systems with large file support disabled by default, you get the 32 bit version of the stat structure. When using the library from programs that change the ABI to use 64 bit file offsets the programs will fail. In that case you have to compile libev with the same flags to get binary compatibility. This is obviously the case with any flags that change the ABI, but the problem is most noticeably displayed with ev_stat and large file support.

The solution for this is to lobby your distribution maker to make large file interfaces available by default (as e.g. FreeBSD does) and not optional. Libev cannot simply switch on large file support because it has to exchange stat structures with application programs compiled using the default compilation environment.

Inotify and Kqueue

When "inotify (7)" support has been compiled into libev and present at runtime, it will be used to speed up change detection where possible. The inotify descriptor will be created lazily when the first "ev_stat" watcher is being started.

Inotify presence does not change the semantics of "ev_stat" watchers except that changes might be detected earlier, and in some cases, to avoid making regular "stat" calls. Even in the presence of inotify support there are many cases where libev has to resort to regular "stat" polling, but as long as kernel 2.6.25 or newer is used (2.6.24 and older have too many bugs), the path exists (i.e. stat succeeds), and the path resides on a local filesystem (libev currently assumes only ext2/3, jfs, reiserfs and xfs are fully working) libev usually gets away without polling.

There is no support for kqueue, as apparently it cannot be used to implement this functionality, due to the requirement of having a file descriptor open on the object at all times, and detecting renames, unlinks etc. is difficult.

"stat ()" is a synchronous operation

Libev doesn't normally do any kind of I/O itself, and so is not blocking the process. The exception are "ev_stat" watchers - those call "stat ()", which is a synchronous operation.

For local paths, this usually doesn't matter: unless the system is very busy or the intervals between stat's are large, a stat call will be fast, as the path data is usually in memory already (except when starting the watcher).

For networked file systems, calling "stat ()" can block an indefinite time due to network issues, and even under good conditions, a stat call often takes multiple milliseconds.

Therefore, it is best to avoid using "ev_stat" watchers on networked paths, although this is fully supported by libev.

The special problem of stat time resolution

The "stat ()" system call only supports full-second resolution portably, and even on systems where the resolution is higher, most file systems still only support whole seconds.

That means that, if the time is the only thing that changes, you can easily miss updates: on the first update, "ev_stat" detects a change and calls your callback, which does something. When there is another update within the same second, "ev_stat" will be unable to detect unless the stat data does change in other ways (e.g. file size).

The solution to this is to delay acting on a change for slightly more than a second (or till slightly after the next full second boundary), using a roughly one-second-delay "ev_timer" (e.g. "ev_timer_set (w, 0., 1.02); ev_timer_again (loop, w)").

The .02 offset is added to work around small timing inconsistencies of some operating systems (where the second counter of the current time might be be delayed. One such system is the Linux kernel, where a call to "gettimeofday" might return a timestamp with a full second later than a subsequent "time" call - if the equivalent of "time ()" is used to update file times then there will be a small window where the kernel uses the previous second to update file times but libev might already execute the timer callback).

Watcher-Specific Functions and Data Members

ev_stat_init (ev_stat *, callback, const char *path, ev_tstamp interval)

ev_stat_set (ev_stat *, const char *path, ev_tstamp interval)

Configures the watcher to wait for status changes of the given "path". The "interval" is a hint on how quickly a change is expected to be detected and should normally be specified as 0 to let libev choose a suitable value. The memory pointed to by "path" must point to the same path for as long as the watcher is active.

The callback will receive an "EV_STAT" event when a change was detected, relative to the attributes at the time the watcher was started (or the last change was detected).

ev_stat_stat (loop, ev_stat *)

Updates the stat buffer immediately with new values. If you change the watched path in your callback, you could call this function to avoid detecting this change (while introducing a race condition if you are not the only one changing the path). Can also be useful simply to find out the new values.

ev_statdata attr [read-only]

The most-recently detected attributes of the file. Although the type is "ev_statdata", this is usually the (or one of the) "struct stat" types suitable for your system, but you can only rely on the POSIX-standardised members to be present. If the "st_nlink" member is 0, then there was some error while "stat"ing the file.

ev_statdata prev [read-only]

The previous attributes of the file. The callback gets invoked whenever "prev" != "attr", or, more precisely, one or more of these members differ: "st_dev", "st_ino", "st_mode", "st_nlink", "st_uid", "st_gid", "st_rdev", "st_size", "st_atime", "st_mtime", "st_ctime".

ev_tstamp interval [read-only]

The specified interval.

const char *path [read-only]

The file system path that is being watched.

Examples

Example: Watch "/etc/passwd" for attribute changes.

   static void
   passwd_cb (struct ev_loop *loop, ev_stat *w, int revents)
   {
     /* /etc/passwd changed in some way */
     if (w->attr.st_nlink)
       {
         printf ("passwd current size  %ld\n", (long)w->attr.st_size);
         printf ("passwd current atime %ld\n", (long)w->attr.st_mtime);
         printf ("passwd current mtime %ld\n", (long)w->attr.st_mtime);
       }
     else
       /* you shalt not abuse printf for puts */
       puts ("wow, /etc/passwd is not there, expect problems. "
             "if this is windows, they already arrived\n");
   }

   ...
   ev_stat passwd;

   ev_stat_init (&passwd, passwd_cb, "/etc/passwd", 0.);
   ev_stat_start (loop, &passwd);

Example: Like above, but additionally use a one-second delay so we do not miss updates (however, frequent updates will delay processing, too, so one might do the work both on "ev_stat" callback invocation and on "ev_timer" callback invocation).

   static ev_stat passwd;
   static ev_timer timer;

   static void
   timer_cb (EV_P_ ev_timer *w, int revents)
   {
     ev_timer_stop (EV_A_ w);

     /* now it's one second after the most recent passwd change */
   }

   static void
   stat_cb (EV_P_ ev_stat *w, int revents)
   {
     /* reset the one-second timer */
     ev_timer_again (EV_A_ &timer);
   }

   ...
   ev_stat_init (&passwd, stat_cb, "/etc/passwd", 0.);
   ev_stat_start (loop, &passwd);
   ev_timer_init (&timer, timer_cb, 0., 1.02);

"ev_idle" - when you've got nothing better to do...

That is, as long as your process is busy handling sockets or timeouts (or even signals, imagine) of the same or higher priority it will not be triggered. But when your process is idle (or only lower-priority watchers are pending), the idle watchers are being called once per event loop iteration - until stopped, that is, or your process receives more events and becomes busy again with higher priority stuff.

The most noteworthy effect is that as long as any idle watchers are active, the process will not block when waiting for new events.

Apart from keeping your process non-blocking (which is a useful effect on its own sometimes), idle watchers are a good place to do "pseudo-background processing", or delay processing stuff to after the event loop has handled all outstanding events.

Abusing an "ev_idle" watcher for its side-effect

As long as there is at least one active idle watcher, libev will never sleep unnecessarily. Or in other words, it will loop as fast as possible. For this to work, the idle watcher doesn't need to be invoked at all - the lowest priority will do.

This mode of operation can be useful together with an "ev_check" watcher, to do something on each event loop iteration - for example to balance load between different connections.

See "Abusing an ev_check watcher for its side-effect" for a longer example.

Watcher-Specific Functions and Data Members

ev_idle_init (ev_idle *, callback)

Initialises and configures the idle watcher - it has no parameters of any kind. There is a "ev_idle_set" macro, but using it is utterly pointless, believe me.

Examples

Example: Dynamically allocate an "ev_idle" watcher, start it, and in the callback, free it. Also, use no error checking, as usual.

   static void
   idle_cb (struct ev_loop *loop, ev_idle *w, int revents)
   {
     // stop the watcher
     ev_idle_stop (loop, w);

     // now we can free it
     free (w);

     // now do something you wanted to do when the program has
     // no longer anything immediate to do.
   }

   ev_idle *idle_watcher = malloc (sizeof (ev_idle));
   ev_idle_init (idle_watcher, idle_cb);
   ev_idle_start (loop, idle_watcher);

"ev_prepare" and "ev_check" - customise your event loop!

You must not call "ev_run" (or similar functions that enter the current event loop) or "ev_loop_fork" from either "ev_prepare" or "ev_check" watchers. Other loops than the current one are fine, however. The rationale behind this is that you do not need to check for recursion in those watchers, i.e. the sequence will always be "ev_prepare", blocking, "ev_check" so if you have one watcher of each kind they will always be called in pairs bracketing the blocking call.

Their main purpose is to integrate other event mechanisms into libev and their use is somewhat advanced. They could be used, for example, to track variable changes, implement your own watchers, integrate net-snmp or a coroutine library and lots more. They are also occasionally useful if you cache some data and want to flush it before blocking (for example, in X programs you might want to do an "XFlush ()" in an "ev_prepare" watcher).

This is done by examining in each prepare call which file descriptors need to be watched by the other library, registering "ev_io" watchers for them and starting an "ev_timer" watcher for any timeouts (many libraries provide exactly this functionality). Then, in the check watcher, you check for any events that occurred (by checking the pending status of all watchers and stopping them) and call back into the library. The I/O and timer callbacks will never actually be called (but must be valid nevertheless, because you never know, you know?).

As another example, the Perl Coro module uses these hooks to integrate coroutines into libev programs, by yielding to other active coroutines during each prepare and only letting the process block if no coroutines are ready to run (it's actually more complicated: it only runs coroutines with priority higher than or equal to the event loop and one coroutine of lower priority, but only once, using idle watchers to keep the event loop from blocking if lower-priority coroutines are active, thus mapping low-priority coroutines to idle/background tasks).

When used for this purpose, it is recommended to give "ev_check" watchers highest ("EV_MAXPRI") priority, to ensure that they are being run before any other watchers after the poll (this doesn't matter for "ev_prepare" watchers).

Also, "ev_check" watchers (and "ev_prepare" watchers, too) should not activate ("feed") events into libev. While libev fully supports this, they might get executed before other "ev_check" watchers did their job. As "ev_check" watchers are often used to embed other (non-libev) event loops those other event loops might be in an unusable state until their "ev_check" watcher ran (always remind yourself to coexist peacefully with others).

Abusing an "ev_check" watcher for its side-effect

"ev_check" (and less often also "ev_prepare") watchers can also be useful because they are called once per event loop iteration. For example, if you want to handle a large number of connections fairly, you normally only do a bit of work for each active connection, and if there is more work to do, you wait for the next event loop iteration, so other connections have a chance of making progress.

Using an "ev_check" watcher is almost enough: it will be called on the next event loop iteration. However, that isn't as soon as possible - without external events, your "ev_check" watcher will not be invoked.

This is where "ev_idle" watchers come in handy - all you need is a single global idle watcher that is active as long as you have one active "ev_check" watcher. The "ev_idle" watcher makes sure the event loop will not sleep, and the "ev_check" watcher makes sure a callback gets invoked. Neither watcher alone can do that.

Watcher-Specific Functions and Data Members

ev_prepare_init (ev_prepare *, callback)

ev_check_init (ev_check *, callback)

Initialises and configures the prepare or check watcher - they have no parameters of any kind. There are "ev_prepare_set" and "ev_check_set" macros, but using them is utterly, utterly, utterly and completely pointless.

Examples

There are a number of principal ways to embed other event loops or modules into libev. Here are some ideas on how to include libadns into libev (there is a Perl module named "EV::ADNS" that does this, which you could use as a working example. Another Perl module named "EV::Glib" embeds a Glib main context into libev, and finally, "Glib::EV" embeds EV into the Glib event loop).

Method 1: Add IO watchers and a timeout watcher in a prepare handler, and in a check watcher, destroy them and call into libadns. What follows is pseudo-code only of course. This requires you to either use a low priority for the check watcher or use "ev_clear_pending" explicitly, as the callbacks for the IO/timeout watchers might not have been called yet.

   static ev_io iow [nfd];
   static ev_timer tw;

   static void
   io_cb (struct ev_loop *loop, ev_io *w, int revents)
   {
   }

   // create io watchers for each fd and a timer before blocking
   static void
   adns_prepare_cb (struct ev_loop *loop, ev_prepare *w, int revents)
   {
     int timeout = 3600000;
     struct pollfd fds [nfd];
     // actual code will need to loop here and realloc etc.
     adns_beforepoll (ads, fds, &nfd, &timeout, timeval_from (ev_time ()));

     /* the callback is illegal, but won't be called as we stop during check */
     ev_timer_init (&tw, 0, timeout * 1e-3, 0.);
     ev_timer_start (loop, &tw);

     // create one ev_io per pollfd
     for (int i = 0; i < nfd; ++i)
       {
         ev_io_init (iow + i, io_cb, fds [i].fd,
           ((fds [i].events & POLLIN ? EV_READ : 0)
            | (fds [i].events & POLLOUT ? EV_WRITE : 0)));

         fds [i].revents = 0;
         ev_io_start (loop, iow + i);
       }
   }

   // stop all watchers after blocking
   static void
   adns_check_cb (struct ev_loop *loop, ev_check *w, int revents)
   {
     ev_timer_stop (loop, &tw);

     for (int i = 0; i < nfd; ++i)
       {
         // set the relevant poll flags
         // could also call adns_processreadable etc. here
         struct pollfd *fd = fds + i;
         int revents = ev_clear_pending (iow + i);
         if (revents & EV_READ ) fd->revents |= fd->events & POLLIN;
         if (revents & EV_WRITE) fd->revents |= fd->events & POLLOUT;

         // now stop the watcher
         ev_io_stop (loop, iow + i);
       }

     adns_afterpoll (adns, fds, nfd, timeval_from (ev_now (loop));
   }

Method 2: This would be just like method 1, but you run "adns_afterpoll" in the prepare watcher and would dispose of the check watcher.

Method 3: If the module to be embedded supports explicit event notification (libadns does), you can also make use of the actual watcher callbacks, and only destroy/create the watchers in the prepare watcher.

   static void
   timer_cb (EV_P_ ev_timer *w, int revents)
   {
     adns_state ads = (adns_state)w->data;
     update_now (EV_A);

     adns_processtimeouts (ads, &tv_now);
   }

   static void
   io_cb (EV_P_ ev_io *w, int revents)
   {
     adns_state ads = (adns_state)w->data;
     update_now (EV_A);

     if (revents & EV_READ ) adns_processreadable  (ads, w->fd, &tv_now);
     if (revents & EV_WRITE) adns_processwriteable (ads, w->fd, &tv_now);
   }

   // do not ever call adns_afterpoll

Method 4: Do not use a prepare or check watcher because the module you want to embed is not flexible enough to support it. Instead, you can override their poll function. The drawback with this solution is that the main loop is now no longer controllable by EV. The "Glib::EV" module uses this approach, effectively embedding EV as a client into the horrible libglib event loop.

   static gint
   event_poll_func (GPollFD *fds, guint nfds, gint timeout)
   {
     int got_events = 0;

     for (n = 0; n < nfds; ++n)
       // create/start io watcher that sets the relevant bits in fds[n] and increment got_events

     if (timeout >= 0)
       // create/start timer

     // poll
     ev_run (EV_A_ 0);

     // stop timer again
     if (timeout >= 0)
       ev_timer_stop (EV_A_ &to);

     // stop io watchers again - their callbacks should have set
     for (n = 0; n < nfds; ++n)
       ev_io_stop (EV_A_ iow [n]);

     return got_events;
   }

"ev_embed" - when one backend isn't enough...

There are primarily two reasons you would want that: work around bugs and prioritise I/O.

As an example for a bug workaround, the kqueue backend might only support sockets on some platform, so it is unusable as generic backend, but you still want to make use of it because you have many sockets and it scales so nicely. In this case, you would create a kqueue-based loop and embed it into your default loop (which might use e.g. poll). Overall operation will be a bit slower because first libev has to call "poll" and then "kevent", but at least you can use both mechanisms for what they are best: "kqueue" for scalable sockets and "poll" if you want it to work :)

As for prioritising I/O: under rare circumstances you have the case where some fds have to be watched and handled very quickly (with low latency), and even priorities and idle watchers might have too much overhead. In this case you would put all the high priority stuff in one loop and all the rest in a second one, and embed the second one in the first.

As long as the watcher is active, the callback will be invoked every time there might be events pending in the embedded loop. The callback must then call "ev_embed_sweep (mainloop, watcher)" to make a single sweep and invoke their callbacks (the callback doesn't need to invoke the "ev_embed_sweep" function directly, it could also start an idle watcher to give the embedded loop strictly lower priority for example).

You can also set the callback to 0, in which case the embed watcher will automatically execute the embedded loop sweep whenever necessary.

Fork detection will be handled transparently while the "ev_embed" watcher is active, i.e., the embedded loop will automatically be forked when the embedding loop forks. In other cases, the user is responsible for calling "ev_loop_fork" on the embedded loop.

Unfortunately, not all backends are embeddable: only the ones returned by "ev_embeddable_backends" are, which, unfortunately, does not include any portable one.

So when you want to use this feature you will always have to be prepared that you cannot get an embeddable loop. The recommended way to get around this is to have a separate variables for your embeddable loop, try to create it, and if that fails, use the normal loop for everything.

"ev_embed" and fork

While the "ev_embed" watcher is running, forks in the embedding loop will automatically be applied to the embedded loop as well, so no special fork handling is required in that case. When the watcher is not running, however, it is still the task of the libev user to call "ev_loop_fork ()" as applicable.

Watcher-Specific Functions and Data Members

ev_embed_init (ev_embed *, callback, struct ev_loop *embedded_loop)

ev_embed_set (ev_embed *, struct ev_loop *embedded_loop)

Configures the watcher to embed the given loop, which must be embeddable. If the callback is 0, then "ev_embed_sweep" will be invoked automatically, otherwise it is the responsibility of the callback to invoke it (it will continue to be called until the sweep has been done, if you do not want that, you need to temporarily stop the embed watcher).

ev_embed_sweep (loop, ev_embed *)

Make a single, non-blocking sweep over the embedded loop. This works similarly to "ev_run (embedded_loop, EVRUN_NOWAIT)", but in the most appropriate way for embedded loops.

struct ev_loop *other [read-only]

The embedded event loop.

Examples

Example: Try to get an embeddable event loop and embed it into the default event loop. If that is not possible, use the default loop. The default loop is stored in "loop_hi", while the embeddable loop is stored in "loop_lo" (which is "loop_hi" in the case no embeddable loop can be used).

   struct ev_loop *loop_hi = ev_default_init (0);
   struct ev_loop *loop_lo = 0;
   ev_embed embed;

   // see if there is a chance of getting one that works
   // (remember that a flags value of 0 means autodetection)
   loop_lo = ev_embeddable_backends () & ev_recommended_backends ()
     ? ev_loop_new (ev_embeddable_backends () & ev_recommended_backends ())
     : 0;

   // if we got one, then embed it, otherwise default to loop_hi
   if (loop_lo)
     {
       ev_embed_init (&embed, 0, loop_lo);
       ev_embed_start (loop_hi, &embed);
     }
   else
     loop_lo = loop_hi;

Example: Check if kqueue is available but not recommended and create a kqueue backend for use with sockets (which usually work with any kqueue implementation). Store the kqueue/socket-only event loop in "loop_socket". (One might optionally use "EVFLAG_NOENV", too).

   struct ev_loop *loop = ev_default_init (0);
   struct ev_loop *loop_socket = 0;
   ev_embed embed;

   if (ev_supported_backends () & ~ev_recommended_backends () & EVBACKEND_KQUEUE)
     if ((loop_socket = ev_loop_new (EVBACKEND_KQUEUE))
       {
         ev_embed_init (&embed, 0, loop_socket);
         ev_embed_start (loop, &embed);
       }

   if (!loop_socket)
     loop_socket = loop;

   // now use loop_socket for all sockets, and loop for everything else

"ev_fork" - the audacity to resume the event loop after a fork

The special problem of life after fork - how is it possible?

Most uses of "fork ()" consist of forking, then some simple calls to set up/change the process environment, followed by a call to "exec()". This sequence should be handled by libev without any problems.

This changes when the application actually wants to do event handling in the child, or both parent in child, in effect "continuing" after the fork.

The default mode of operation (for libev, with application help to detect forks) is to duplicate all the state in the child, as would be expected when either the parent or the child process continues.

When both processes want to continue using libev, then this is usually the wrong result. In that case, usually one process (typically the parent) is supposed to continue with all watchers in place as before, while the other process typically wants to start fresh, i.e. without any active watchers.

The cleanest and most efficient way to achieve that with libev is to simply create a new event loop, which of course will be "empty", and use that for new watchers. This has the advantage of not touching more memory than necessary, and thus avoiding the copy-on-write, and the disadvantage of having to use multiple event loops (which do not support signal watchers).

When this is not possible, or you want to use the default loop for other reasons, then in the process that wants to start "fresh", call "ev_loop_destroy (EV_DEFAULT)" followed by "ev_default_loop (...)". Destroying the default loop will "orphan" (not stop) all registered watchers, so you have to be careful not to execute code that modifies those watchers. Note also that in that case, you have to re-register any signal watchers.

Watcher-Specific Functions and Data Members

ev_fork_init (ev_fork *, callback)

Initialises and configures the fork watcher - it has no parameters of any kind. There is a "ev_fork_set" macro, but using it is utterly pointless, really.

"ev_cleanup" - even the best things end

While there is no guarantee that the event loop gets destroyed, cleanup watchers provide a convenient method to install cleanup hooks for your program, worker threads and so on - you just to make sure to destroy the loop when you want them to be invoked.

Cleanup watchers are invoked in the same way as any other watcher. Unlike all other watchers, they do not keep a reference to the event loop (which makes a lot of sense if you think about it). Like all other watchers, you can call libev functions in the callback, except "ev_cleanup_start".

Watcher-Specific Functions and Data Members

ev_cleanup_init (ev_cleanup *, callback)

Initialises and configures the cleanup watcher - it has no parameters of any kind. There is a "ev_cleanup_set" macro, but using it is utterly pointless, I assure you.

Example: Register an atexit handler to destroy the default loop, so any cleanup functions are called.

   static void
   program_exits (void)
   {
     ev_loop_destroy (EV_DEFAULT_UC);
   }

   ...
   atexit (program_exits);

"ev_async" - how to wake up an event loop

Sometimes, however, you need to wake up an event loop you do not control, for example because it belongs to another thread. This is what "ev_async" watchers do: as long as the "ev_async" watcher is active, you can signal it by calling "ev_async_send", which is thread- and signal safe.

This functionality is very similar to "ev_signal" watchers, as signals, too, are asynchronous in nature, and signals, too, will be compressed (i.e. the number of callback invocations may be less than the number of "ev_async_send" calls). In fact, you could use signal watchers as a kind of "global async watchers" by using a watcher on an otherwise unused signal, and "ev_feed_signal" to signal this watcher from another thread, even without knowing which loop owns the signal.

Queueing

"ev_async" does not support queueing of data in any way. The reason is that the author does not know of a simple (or any) algorithm for a multiple-writer-single-reader queue that works in all cases and doesn't need elaborate support such as pthreads or unportable memory access semantics.

That means that if you want to queue data, you have to provide your own queue. But at least I can tell you how to implement locking around your queue:

queueing from a signal handler context

To implement race-free queueing, you simply add to the queue in the signal handler but you block the signal handler in the watcher callback. Here is an example that does that for some fictitious SIGUSR1 handler:

   static ev_async mysig;

   static void
   sigusr1_handler (void)
   {
     sometype data;

     // no locking etc.
     queue_put (data);
     ev_async_send (EV_DEFAULT_ &mysig);
   }

   static void
   mysig_cb (EV_P_ ev_async *w, int revents)
   {
     sometype data;
     sigset_t block, prev;

     sigemptyset (&block);
     sigaddset (&block, SIGUSR1);
     sigprocmask (SIG_BLOCK, &block, &prev);

     while (queue_get (&data))
       process (data);

     if (sigismember (&prev, SIGUSR1)
       sigprocmask (SIG_UNBLOCK, &block, 0);
   }
    

(Note: pthreads in theory requires you to use "pthread_setmask" instead of "sigprocmask" when you use threads, but libev doesn't do it either...).

queueing from a thread context

The strategy for threads is different, as you cannot (easily) block threads but you can easily preempt them, so to queue safely you need to employ a traditional mutex lock, such as in this pthread example:

   static ev_async mysig;
   static pthread_mutex_t mymutex = PTHREAD_MUTEX_INITIALIZER;

   static void
   otherthread (void)
   {
     // only need to lock the actual queueing operation
     pthread_mutex_lock (&mymutex);
     queue_put (data);
     pthread_mutex_unlock (&mymutex);

     ev_async_send (EV_DEFAULT_ &mysig);
   }

   static void
   mysig_cb (EV_P_ ev_async *w, int revents)
   {
     pthread_mutex_lock (&mymutex);

     while (queue_get (&data))
       process (data);

     pthread_mutex_unlock (&mymutex);
   }
    

Watcher-Specific Functions and Data Members

ev_async_init (ev_async *, callback)

Initialises and configures the async watcher - it has no parameters of any kind. There is a "ev_async_set" macro, but using it is utterly pointless, trust me.

ev_async_send (loop, ev_async *)

Sends/signals/activates the given "ev_async" watcher, that is, feeds an "EV_ASYNC" event on the watcher into the event loop, and instantly returns.

Unlike "ev_feed_event", this call is safe to do from other threads, signal or similar contexts (see the discussion of "EV_ATOMIC_T" in the embedding section below on what exactly this means).

Note that, as with other watchers in libev, multiple events might get compressed into a single callback invocation (another way to look at this is that "ev_async" watchers are level-triggered: they are set on "ev_async_send", reset when the event loop detects that).

This call incurs the overhead of at most one extra system call per event loop iteration, if the event loop is blocked, and no syscall at all if the event loop (or your program) is processing events. That means that repeated calls are basically free (there is no need to avoid calls for performance reasons) and that the overhead becomes smaller (typically zero) under load.

bool = ev_async_pending (ev_async *)

Returns a non-zero value when "ev_async_send" has been called on the watcher but the event has not yet been processed (or even noted) by the event loop.

"ev_async_send" sets a flag in the watcher and wakes up the loop. When the loop iterates next and checks for the watcher to have become active, it will reset the flag again. "ev_async_pending" can be used to very quickly check whether invoking the loop might be a good idea.

Not that this does not check whether the watcher itself is pending, only whether it has been requested to make this watcher pending: there is a time window between the event loop checking and resetting the async notification, and the callback being invoked.

ASSOCIATING CUSTOM DATA WITH A WATCHER

   struct my_io
   {
     ev_io io;
     int otherfd;
     void *somedata;
     struct whatever *mostinteresting;
   };

   ...
   struct my_io w;
   ev_io_init (&w.io, my_cb, fd, EV_READ);

And since your callback will be called with a pointer to the watcher, you can cast it back to your own type:

   static void my_cb (struct ev_loop *loop, ev_io *w_, int revents)
   {
     struct my_io *w = (struct my_io *)w_;
     ...
   }

More interesting and less C-conformant ways of casting your callback function type instead have been omitted.

BUILDING YOUR OWN COMPOSITE WATCHERS

   struct my_biggy
   {
     int some_data;
     ev_timer t1;
     ev_timer t2;
   }

In this case getting the pointer to "my_biggy" is a bit more complicated: Either you store the address of your "my_biggy" struct in the "data" member of the watcher (for woozies or C++ coders), or you need to use some pointer arithmetic using "offsetof" inside your watchers (for real programmers):

   #include <stddef.h>

   static void
   t1_cb (EV_P_ ev_timer *w, int revents)
   {
     struct my_biggy big = (struct my_biggy *)
       (((char *)w) - offsetof (struct my_biggy, t1));
   }

   static void
   t2_cb (EV_P_ ev_timer *w, int revents)
   {
     struct my_biggy big = (struct my_biggy *)
       (((char *)w) - offsetof (struct my_biggy, t2));
   }

AVOIDING FINISHING BEFORE RETURNING

  callback ()
  {
    free (request);
  }

  request = start_new_request (..., callback);

The intent is to start some "lengthy" operation. The "request" could be used to cancel the operation, or do other things with it.

It's not uncommon to have code paths in "start_new_request" that immediately invoke the callback, for example, to report errors. Or you add some caching layer that finds that it can skip the lengthy aspects of the operation and simply invoke the callback with the result.

The problem here is that this will happen before "start_new_request" has returned, so "request" is not set.

Even if you pass the request by some safer means to the callback, you might want to do something to the request after starting it, such as canceling it, which probably isn't working so well when the callback has already been invoked.

A common way around all these issues is to make sure that "start_new_request" always returns before the callback is invoked. If "start_new_request" immediately knows the result, it can artificially delay invoking the callback by using a "prepare" or "idle" watcher for example, or more sneakily, by reusing an existing (stopped) watcher and pushing it into the pending queue:

   ev_set_cb (watcher, callback);
   ev_feed_event (EV_A_ watcher, 0);

This way, "start_new_request" can safely return before the callback is invoked, while not delaying callback invocation too much.

MODEL/NESTED EVENT LOOP INVOCATIONS AND EXIT CONDITIONS

This brings the problem of exiting - a callback might want to finish the main "ev_run" call, but not the nested one (e.g. user clicked "Quit", but a modal "Are you sure?" dialog is still waiting), or just the nested one and not the main one (e.g. user clocked "Ok" in a modal dialog), or some other combination: In these cases, a simple "ev_break" will not work.

The solution is to maintain "break this loop" variable for each "ev_run" invocation, and use a loop around "ev_run" until the condition is triggered, using "EVRUN_ONCE":

   // main loop
   int exit_main_loop = 0;

   while (!exit_main_loop)
     ev_run (EV_DEFAULT_ EVRUN_ONCE);

   // in a modal watcher
   int exit_nested_loop = 0;

   while (!exit_nested_loop)
     ev_run (EV_A_ EVRUN_ONCE);

To exit from any of these loops, just set the corresponding exit variable:

   // exit modal loop
   exit_nested_loop = 1;

   // exit main program, after modal loop is finished
   exit_main_loop = 1;

   // exit both
   exit_main_loop = exit_nested_loop = 1;

THREAD LOCKING EXAMPLE

For a real-world example, see the "EV::Loop::Async" perl module, which uses exactly this technique (which is suited for many high-level languages).

The example uses a pthread mutex to protect the loop data, a condition variable to wait for callback invocations, an async watcher to notify the event loop thread and an unspecified mechanism to wake up the main thread.

First, you need to associate some data with the event loop:

   typedef struct {
     mutex_t lock; /* global loop lock */
     ev_async async_w;
     thread_t tid;
     cond_t invoke_cv;
   } userdata;

   void prepare_loop (EV_P)
   {
      // for simplicity, we use a static userdata struct.
      static userdata u;

      ev_async_init (&u->async_w, async_cb);
      ev_async_start (EV_A_ &u->async_w);

      pthread_mutex_init (&u->lock, 0);
      pthread_cond_init (&u->invoke_cv, 0);

      // now associate this with the loop
      ev_set_userdata (EV_A_ u);
      ev_set_invoke_pending_cb (EV_A_ l_invoke);
      ev_set_loop_release_cb (EV_A_ l_release, l_acquire);

      // then create the thread running ev_run
      pthread_create (&u->tid, 0, l_run, EV_A);
   }

The callback for the "ev_async" watcher does nothing: the watcher is used solely to wake up the event loop so it takes notice of any new watchers that might have been added:

   static void
   async_cb (EV_P_ ev_async *w, int revents)
   {
      // just used for the side effects
   }

The "l_release" and "l_acquire" callbacks simply unlock/lock the mutex protecting the loop data, respectively.

   static void
   l_release (EV_P)
   {
     userdata *u = ev_userdata (EV_A);
     pthread_mutex_unlock (&u->lock);
   }

   static void
   l_acquire (EV_P)
   {
     userdata *u = ev_userdata (EV_A);
     pthread_mutex_lock (&u->lock);
   }

The event loop thread first acquires the mutex, and then jumps straight into "ev_run":

   void *
   l_run (void *thr_arg)
   {
     struct ev_loop *loop = (struct ev_loop *)thr_arg;

     l_acquire (EV_A);
     pthread_setcanceltype (PTHREAD_CANCEL_ASYNCHRONOUS, 0);
     ev_run (EV_A_ 0);
     l_release (EV_A);

     return 0;
   }

Instead of invoking all pending watchers, the "l_invoke" callback will signal the main thread via some unspecified mechanism (signals? pipe writes? "Async::Interrupt"?) and then waits until all pending watchers have been called (in a while loop because a) spurious wakeups are possible and b) skipping inter-thread-communication when there are no pending watchers is very beneficial):

   static void
   l_invoke (EV_P)
   {
     userdata *u = ev_userdata (EV_A);

     while (ev_pending_count (EV_A))
       {
         wake_up_other_thread_in_some_magic_or_not_so_magic_way ();
         pthread_cond_wait (&u->invoke_cv, &u->lock);
       }
   }

Now, whenever the main thread gets told to invoke pending watchers, it will grab the lock, call "ev_invoke_pending" and then signal the loop thread to continue:

   static void
   real_invoke_pending (EV_P)
   {
     userdata *u = ev_userdata (EV_A);

     pthread_mutex_lock (&u->lock);
     ev_invoke_pending (EV_A);
     pthread_cond_signal (&u->invoke_cv);
     pthread_mutex_unlock (&u->lock);
   }

Whenever you want to start/stop a watcher or do other modifications to an event loop, you will now have to lock:

   ev_timer timeout_watcher;
   userdata *u = ev_userdata (EV_A);

   ev_timer_init (&timeout_watcher, timeout_cb, 5.5, 0.);

   pthread_mutex_lock (&u->lock);
   ev_timer_start (EV_A_ &timeout_watcher);
   ev_async_send (EV_A_ &u->async_w);
   pthread_mutex_unlock (&u->lock);

Note that sending the "ev_async" watcher is required because otherwise an event loop currently blocking in the kernel will have no knowledge about the newly added timer. By waking up the loop it will pick up any new watchers in the next event loop iteration.

THREADS, COROUTINES, CONTINUATIONS, QUEUES... INSTEAD OF CALLBACKS

Imagine you have coroutines that you can switch to using a function "switch_to (coro)", that libev runs in a coroutine called "libev_coro" and that due to some magic, the currently active coroutine is stored in a global called "current_coro". Then you can build your own "wait for libev event" primitive by changing "EV_CB_DECLARE" and "EV_CB_INVOKE" (note the differing ";" conventions):

   #define EV_CB_DECLARE(type)   struct my_coro *cb;
   #define EV_CB_INVOKE(watcher) switch_to ((watcher)->cb)

That means instead of having a C callback function, you store the coroutine to switch to in each watcher, and instead of having libev call your callback, you instead have it switch to that coroutine.

A coroutine might now wait for an event with a function called "wait_for_event". (the watcher needs to be started, as always, but it doesn't matter when, or whether the watcher is active or not when this function is called):

   void
   wait_for_event (ev_watcher *w)
   {
     ev_set_cb (w, current_coro);
     switch_to (libev_coro);
   }

That basically suspends the coroutine inside "wait_for_event" and continues the libev coroutine, which, when appropriate, switches back to this or any other coroutine.

You can do similar tricks if you have, say, threads with an event queue - instead of storing a coroutine, you store the queue object and instead of switching to a coroutine, you push the watcher onto the queue and notify any waiters.

To embed libev, see "EMBEDDING", but in short, it's easiest to create two files, my_ev.h and my_ev.c that include the respective libev files:

   // my_ev.h
   #define EV_CB_DECLARE(type)   struct my_coro *cb;
   #define EV_CB_INVOKE(watcher) switch_to ((watcher)->cb)
   #include "../libev/ev.h"

   // my_ev.c
   #define EV_H "my_ev.h"
   #include "../libev/ev.c"

And then use my_ev.h when you would normally use ev.h, and compile my_ev.c into your project. When properly specifying include paths, you can even use ev.h as header file name directly.

C API

Proper exception specifications might have to be added to callbacks passed to libev: exceptions may be thrown only from watcher callbacks, all other callbacks (allocator, syserr, loop acquire/release and periodic reschedule callbacks) must not throw exceptions, and might need a "noexcept" specification. If you have code that needs to be compiled as both C and C++ you can use the "EV_NOEXCEPT" macro for this:

   static void
   fatal_error (const char *msg) EV_NOEXCEPT
   {
     perror (msg);
     abort ();
   }

   ...
   ev_set_syserr_cb (fatal_error);

The only API functions that can currently throw exceptions are "ev_run", "ev_invoke", "ev_invoke_pending" and "ev_loop_destroy" (the latter because it runs cleanup watchers).

Throwing exceptions in watcher callbacks is only supported if libev itself is compiled with a C++ compiler or your C and C++ environments allow throwing exceptions through C libraries (most do).

C++ API

To use it,

   #include <ev++.h>

This automatically includes ev.h and puts all of its definitions (many of them macros) into the global namespace. All C++ specific things are put into the "ev" namespace. It should support all the same embedding options as ev.h, most notably "EV_MULTIPLICITY".

Care has been taken to keep the overhead low. The only data member the C++ classes add (compared to plain C-style watchers) is the event loop pointer that the watcher is associated with (or no additional members at all if you disable "EV_MULTIPLICITY" when embedding libev).

Currently, functions, static and non-static member functions and classes with "operator ()" can be used as callbacks. Other types should be easy to add as long as they only need one additional pointer for context. If you need support for other types of functors please contact the author (preferably after implementing it).

For all this to work, your C++ compiler either has to use the same calling conventions as your C compiler (for static member functions), or you have to embed libev and compile libev itself as C++.

Here is a list of things available in the "ev" namespace:

"ev::READ", "ev::WRITE" etc.

These are just enum values with the same values as the "EV_READ" etc. macros from ev.h.

"ev::tstamp", "ev::now"

Aliases to the same types/functions as with the "ev_" prefix.

"ev::io", "ev::timer", "ev::periodic", "ev::idle", "ev::sig" etc.

For each "ev_TYPE" watcher in ev.h there is a corresponding class of the same name in the "ev" namespace, with the exception of "ev_signal" which is called "ev::sig" to avoid clashes with the "signal" macro defined by many implementations.

All of those classes have these methods:

Example: Define a class with two I/O and idle watchers, start the I/O watchers in the constructor.

   class myclass
   {
     ev::io   io  ; void io_cb   (ev::io   &w, int revents);
     ev::io   io2 ; void io2_cb  (ev::io   &w, int revents);
     ev::idle idle; void idle_cb (ev::idle &w, int revents);

     myclass (int fd)
     {
       io  .set <myclass, &myclass::io_cb  > (this);
       io2 .set <myclass, &myclass::io2_cb > (this);
       idle.set <myclass, &myclass::idle_cb> (this);

       io.set (fd, ev::WRITE); // configure the watcher
       io.start ();            // start it whenever convenient

       io2.start (fd, ev::READ); // set + start in one call
     }
   };

FILESETS

CORE EVENT LOOP

To include only the libev core (all the "ev_*" functions), with manual configuration (no autoconf):

   #define EV_STANDALONE 1
   #include "ev.c"

This will automatically include ev.h, too, and should be done in a single C source file only to provide the function implementations. To use it, do the same for ev.h in all files wishing to use this API (best done by writing a wrapper around ev.h that you can include instead and where you can put other configuration options):

   #define EV_STANDALONE 1
   #include "ev.h"

Both header files and implementation files can be compiled with a C++ compiler (at least, that's a stated goal, and breakage will be treated as a bug).

You need the following files in your source tree, or in a directory in your include path (e.g. in libev/ when using -Ilibev):

   ev.h
   ev.c
   ev_vars.h
   ev_wrap.h

   ev_win32.c      required on win32 platforms only

   ev_select.c     only when select backend is enabled
   ev_poll.c       only when poll backend is enabled
   ev_epoll.c      only when the epoll backend is enabled
   ev_linuxaio.c   only when the linux aio backend is enabled
   ev_iouring.c    only when the linux io_uring backend is enabled
   ev_kqueue.c     only when the kqueue backend is enabled
   ev_port.c       only when the solaris port backend is enabled

ev.c includes the backend files directly when enabled, so you only need to compile this single file.

LIBEVENT COMPATIBILITY API

To include the libevent compatibility API, also include:

   #include "event.c"

in the file including ev.c, and:

   #include "event.h"

in the files that want to use the libevent API. This also includes ev.h.

You need the following additional files for this:

   event.h
   event.c

AUTOCONF SUPPORT

Instead of using "EV_STANDALONE=1" and providing your configuration in whatever way you want, you can also "m4_include([libev.m4])" in your configure.ac and leave "EV_STANDALONE" undefined. ev.c will then include config.h and configure itself accordingly.

For this of course you need the m4 file:

   libev.m4

PREPROCESSOR SYMBOLS/MACROS

Symbols marked with "(h)" do not change the ABI, and can have different values when compiling libev vs. including ev.h, so it is permissible to redefine them before including ev.h without breaking compatibility to a compiled library. All other symbols change the ABI, which means all users of libev and the libev code itself must be compiled with compatible settings.

EV_COMPAT3 (h)

Backwards compatibility is a major concern for libev. This is why this release of libev comes with wrappers for the functions and symbols that have been renamed between libev version 3 and 4.

You can disable these wrappers (to test compatibility with future versions) by defining "EV_COMPAT3" to 0 when compiling your sources. This has the additional advantage that you can drop the "struct" from "struct ev_loop" declarations, as libev will provide an "ev_loop" typedef in that case.

In some future version, the default for "EV_COMPAT3" will become 0, and in some even more future version the compatibility code will be removed completely.

EV_STANDALONE (h)

Must always be 1 if you do not use autoconf configuration, which keeps libev from including config.h, and it also defines dummy implementations for some libevent functions (such as logging, which is not supported). It will also not define any of the structs usually found in event.h that are not directly supported by the libev core alone.

In standalone mode, libev will still try to automatically deduce the configuration, but has to be more conservative.

EV_USE_FLOOR

If defined to be 1, libev will use the "floor ()" function for its periodic reschedule calculations, otherwise libev will fall back on a portable (slower) implementation. If you enable this, you usually have to link against libm or something equivalent. Enabling this when the "floor" function is not available will fail, so the safe default is to not enable this.

EV_USE_MONOTONIC

If defined to be 1, libev will try to detect the availability of the monotonic clock option at both compile time and runtime. Otherwise no use of the monotonic clock option will be attempted. If you enable this, you usually have to link against librt or something similar. Enabling it when the functionality isn't available is safe, though, although you have to make sure you link against any libraries where the "clock_gettime" function is hiding in (often -lrt). See also "EV_USE_CLOCK_SYSCALL".

EV_USE_REALTIME

If defined to be 1, libev will try to detect the availability of the real-time clock option at compile time (and assume its availability at runtime if successful). Otherwise no use of the real-time clock option will be attempted. This effectively replaces "gettimeofday" by "clock_get (CLOCK_REALTIME, ...)" and will not normally affect correctness. See the note about libraries in the description of "EV_USE_MONOTONIC", though. Defaults to the opposite value of "EV_USE_CLOCK_SYSCALL".

EV_USE_CLOCK_SYSCALL

If defined to be 1, libev will try to use a direct syscall instead of calling the system-provided "clock_gettime" function. This option exists because on GNU/Linux, "clock_gettime" is in "librt", but "librt" unconditionally pulls in "libpthread", slowing down single-threaded programs needlessly. Using a direct syscall is slightly slower (in theory), because no optimised vdso implementation can be used, but avoids the pthread dependency. Defaults to 1 on GNU/Linux with glibc 2.x or higher, as it simplifies linking (no need for "-lrt").

EV_USE_NANOSLEEP

If defined to be 1, libev will assume that "nanosleep ()" is available and will use it for delays. Otherwise it will use "select ()".

EV_USE_EVENTFD

If defined to be 1, then libev will assume that "eventfd ()" is available and will probe for kernel support at runtime. This will improve "ev_signal" and "ev_async" performance and reduce resource consumption. If undefined, it will be enabled if the headers indicate GNU/Linux + Glibc 2.7 or newer, otherwise disabled.

EV_USE_SIGNALFD

If defined to be 1, then libev will assume that "signalfd ()" is available and will probe for kernel support at runtime. This enables the use of EVFLAG_SIGNALFD for faster and simpler signal handling. If undefined, it will be enabled if the headers indicate GNU/Linux + Glibc 2.7 or newer, otherwise disabled.

EV_USE_TIMERFD

If defined to be 1, then libev will assume that "timerfd ()" is available and will probe for kernel support at runtime. This allows libev to detect time jumps accurately. If undefined, it will be enabled if the headers indicate GNU/Linux + Glibc 2.8 or newer and define "TFD_TIMER_CANCEL_ON_SET", otherwise disabled.

EV_USE_EVENTFD

If defined to be 1, then libev will assume that "eventfd ()" is available and will probe for kernel support at runtime. This will improve "ev_signal" and "ev_async" performance and reduce resource consumption. If undefined, it will be enabled if the headers indicate GNU/Linux + Glibc 2.7 or newer, otherwise disabled.

EV_USE_SELECT

If undefined or defined to be 1, libev will compile in support for the "select"(2) backend. No attempt at auto-detection will be done: if no other method takes over, select will be it. Otherwise the select backend will not be compiled in.

EV_SELECT_USE_FD_SET

If defined to 1, then the select backend will use the system "fd_set" structure. This is useful if libev doesn't compile due to a missing "NFDBITS" or "fd_mask" definition or it mis-guesses the bitset layout on exotic systems. This usually limits the range of file descriptors to some low limit such as 1024 or might have other limitations (winsocket only allows 64 sockets). The "FD_SETSIZE" macro, set before compilation, configures the maximum size of the "fd_set".

EV_SELECT_IS_WINSOCKET

When defined to 1, the select backend will assume that select/socket/connect etc. don't understand file descriptors but wants osf handles on win32 (this is the case when the select to be used is the winsock select). This means that it will call "_get_osfhandle" on the fd to convert it to an OS handle. Otherwise, it is assumed that all these functions actually work on fds, even on win32. Should not be defined on non-win32 platforms.

EV_FD_TO_WIN32_HANDLE(fd)

If "EV_SELECT_IS_WINSOCKET" is enabled, then libev needs a way to map file descriptors to socket handles. When not defining this symbol (the default), then libev will call "_get_osfhandle", which is usually correct. In some cases, programs use their own file descriptor management, in which case they can provide this function to map fds to socket handles.

EV_WIN32_HANDLE_TO_FD(handle)

If "EV_SELECT_IS_WINSOCKET" then libev maps handles to file descriptors using the standard "_open_osfhandle" function. For programs implementing their own fd to handle mapping, overwriting this function makes it easier to do so. This can be done by defining this macro to an appropriate value.

EV_WIN32_CLOSE_FD(fd)

If programs implement their own fd to handle mapping on win32, then this macro can be used to override the "close" function, useful to unregister file descriptors again. Note that the replacement function has to close the underlying OS handle.

EV_USE_WSASOCKET

If defined to be 1, libev will use "WSASocket" to create its internal communication socket, which works better in some environments. Otherwise, the normal "socket" function will be used, which works better in other environments.

EV_USE_POLL

If defined to be 1, libev will compile in support for the "poll"(2) backend. Otherwise it will be enabled on non-win32 platforms. It takes precedence over select.

EV_USE_EPOLL

If defined to be 1, libev will compile in support for the Linux "epoll"(7) backend. Its availability will be detected at runtime, otherwise another method will be used as fallback. This is the preferred backend for GNU/Linux systems. If undefined, it will be enabled if the headers indicate GNU/Linux + Glibc 2.4 or newer, otherwise disabled.

EV_USE_LINUXAIO

If defined to be 1, libev will compile in support for the Linux aio backend ("EV_USE_EPOLL" must also be enabled). If undefined, it will be enabled on linux, otherwise disabled.

EV_USE_IOURING

If defined to be 1, libev will compile in support for the Linux io_uring backend ("EV_USE_EPOLL" must also be enabled). Due to it's current limitations it has to be requested explicitly. If undefined, it will be enabled on linux, otherwise disabled.

EV_USE_KQUEUE

If defined to be 1, libev will compile in support for the BSD style "kqueue"(2) backend. Its actual availability will be detected at runtime, otherwise another method will be used as fallback. This is the preferred backend for BSD and BSD-like systems, although on most BSDs kqueue only supports some types of fds correctly (the only platform we found that supports ptys for example was NetBSD), so kqueue might be compiled in, but not be used unless explicitly requested. The best way to use it is to find out whether kqueue supports your type of fd properly and use an embedded kqueue loop.

EV_USE_PORT

If defined to be 1, libev will compile in support for the Solaris 10 port style backend. Its availability will be detected at runtime, otherwise another method will be used as fallback. This is the preferred backend for Solaris 10 systems.

EV_USE_DEVPOLL

Reserved for future expansion, works like the USE symbols above.

EV_USE_INOTIFY

If defined to be 1, libev will compile in support for the Linux inotify interface to speed up "ev_stat" watchers. Its actual availability will be detected at runtime. If undefined, it will be enabled if the headers indicate GNU/Linux + Glibc 2.4 or newer, otherwise disabled.

EV_NO_SMP

If defined to be 1, libev will assume that memory is always coherent between threads, that is, threads can be used, but threads never run on different cpus (or different cpu cores). This reduces dependencies and makes libev faster.

EV_NO_THREADS

If defined to be 1, libev will assume that it will never be called from different threads (that includes signal handlers), which is a stronger assumption than "EV_NO_SMP", above. This reduces dependencies and makes libev faster.

EV_ATOMIC_T

Libev requires an integer type (suitable for storing 0 or 1) whose access is atomic with respect to other threads or signal contexts. No such type is easily found in the C language, so you can provide your own type that you know is safe for your purposes. It is used both for signal handler "locking" as well as for signal and thread safety in "ev_async" watchers.

In the absence of this define, libev will use "sig_atomic_t volatile" (from signal.h), which is usually good enough on most platforms.

EV_H (h)

The name of the ev.h header file used to include it. The default if undefined is "ev.h" in event.h, ev.c and ev++.h. This can be used to virtually rename the ev.h header file in case of conflicts.

EV_CONFIG_H (h)

If "EV_STANDALONE" isn't 1, this variable can be used to override ev.c's idea of where to find the config.h file, similarly to "EV_H", above.

EV_EVENT_H (h)

Similarly to "EV_H", this macro can be used to override event.c's idea of how the event.h header can be found, the default is "event.h".

EV_PROTOTYPES (h)

If defined to be 0, then ev.h will not define any function prototypes, but still define all the structs and other symbols. This is occasionally useful if you want to provide your own wrapper functions around libev functions.

EV_MULTIPLICITY

If undefined or defined to 1, then all event-loop-specific functions will have the "struct ev_loop *" as first argument, and you can create additional independent event loops. Otherwise there will be no support for multiple event loops and there is no first event loop pointer argument. Instead, all functions act on the single default loop.

Note that "EV_DEFAULT" and "EV_DEFAULT_" will no longer provide a default loop when multiplicity is switched off - you always have to initialise the loop manually in this case.

EV_MINPRI

EV_MAXPRI

The range of allowed priorities. "EV_MINPRI" must be smaller or equal to "EV_MAXPRI", but otherwise there are no non-obvious limitations. You can provide for more priorities by overriding those symbols (usually defined to be "-2" and 2, respectively).

When doing priority-based operations, libev usually has to linearly search all the priorities, so having many of them (hundreds) uses a lot of space and time, so using the defaults of five priorities (-2 .. +2) is usually fine.

If your embedding application does not need any priorities, defining these both to 0 will save some memory and CPU.

EV_PERIODIC_ENABLE, EV_IDLE_ENABLE, EV_EMBED_ENABLE, EV_STAT_ENABLE, EV_PREPARE_ENABLE, EV_CHECK_ENABLE, EV_FORK_ENABLE, EV_SIGNAL_ENABLE, EV_ASYNC_ENABLE, EV_CHILD_ENABLE.

If undefined or defined to be 1 (and the platform supports it), then the respective watcher type is supported. If defined to be 0, then it is not. Disabling watcher types mainly saves code size.

EV_FEATURES

If you need to shave off some kilobytes of code at the expense of some speed (but with the full API), you can define this symbol to request certain subsets of functionality. The default is to enable all features that can be enabled on the platform.

A typical way to use this symbol is to define it to 0 (or to a bitset with some broad features you want) and then selectively re-enable additional parts you want, for example if you want everything minimal, but multiple event loop support, async and child watchers and the poll backend, use this:

   #define EV_FEATURES 0
   #define EV_MULTIPLICITY 1
   #define EV_USE_POLL 1
   #define EV_CHILD_ENABLE 1
   #define EV_ASYNC_ENABLE 1
    

The actual value is a bitset, it can be a combination of the following values (by default, all of these are enabled):

EV_API_STATIC

If this symbol is defined (by default it is not), then all identifiers will have static linkage. This means that libev will not export any identifiers, and you cannot link against libev anymore. This can be useful when you embed libev, only want to use libev functions in a single file, and do not want its identifiers to be visible.

To use this, define "EV_API_STATIC" and include ev.c in the file that wants to use libev.

This option only works when libev is compiled with a C compiler, as C++ doesn't support the required declaration syntax.

EV_AVOID_STDIO

If this is set to 1 at compiletime, then libev will avoid using stdio functions (printf, scanf, perror etc.). This will increase the code size somewhat, but if your program doesn't otherwise depend on stdio and your libc allows it, this avoids linking in the stdio library which is quite big.

Note that error messages might become less precise when this option is enabled.

EV_NSIG

The highest supported signal number, +1 (or, the number of signals): Normally, libev tries to deduce the maximum number of signals automatically, but sometimes this fails, in which case it can be specified. Also, using a lower number than detected (32 should be good for about any system in existence) can save some memory, as libev statically allocates some 12-24 bytes per signal number.

EV_PID_HASHSIZE

"ev_child" watchers use a small hash table to distribute workload by pid. The default size is 16 (or 1 with "EV_FEATURES" disabled), usually more than enough. If you need to manage thousands of children you might want to increase this value (must be a power of two).

EV_INOTIFY_HASHSIZE

"ev_stat" watchers use a small hash table to distribute workload by inotify watch id. The default size is 16 (or 1 with "EV_FEATURES" disabled), usually more than enough. If you need to manage thousands of "ev_stat" watchers you might want to increase this value (must be a power of two).

EV_USE_4HEAP

Heaps are not very cache-efficient. To improve the cache-efficiency of the timer and periodics heaps, libev uses a 4-heap when this symbol is defined to 1. The 4-heap uses more complicated (longer) code but has noticeably faster performance with many (thousands) of watchers.

The default is 1, unless "EV_FEATURES" overrides it, in which case it will be 0.

EV_HEAP_CACHE_AT

Heaps are not very cache-efficient. To improve the cache-efficiency of the timer and periodics heaps, libev can cache the timestamp (at) within the heap structure (selected by defining "EV_HEAP_CACHE_AT" to 1), which uses 8-12 bytes more per watcher and a few hundred bytes more code, but avoids random read accesses on heap changes. This improves performance noticeably with many (hundreds) of watchers.

The default is 1, unless "EV_FEATURES" overrides it, in which case it will be 0.

EV_VERIFY

Controls how much internal verification (see "ev_verify ()") will be done: If set to 0, no internal verification code will be compiled in. If set to 1, then verification code will be compiled in, but not called. If set to 2, then the internal verification code will be called once per loop, which can slow down libev. If set to 3, then the verification code will be called very frequently, which will slow down libev considerably.

Verification errors are reported via C's "assert" mechanism, so if you disable that (e.g. by defining "NDEBUG") then no errors will be reported.

The default is 1, unless "EV_FEATURES" overrides it, in which case it will be 0.

EV_COMMON

By default, all watchers have a "void *data" member. By redefining this macro to something else you can include more and other types of members. You have to define it each time you include one of the files, though, and it must be identical each time.

For example, the perl EV module uses something like this:

   #define EV_COMMON                       \
     SV *self; /* contains this struct */  \
     SV *cb_sv, *fh /* note no trailing ";" */
    

EV_CB_DECLARE (type)

EV_CB_INVOKE (watcher, revents)

ev_set_cb (ev, cb)

Can be used to change the callback member declaration in each watcher, and the way callbacks are invoked and set. Must expand to a struct member definition and a statement, respectively. See the ev.h header file for their default definitions. One possible use for overriding these is to avoid the "struct ev_loop *" as first argument in all cases, or to use method calls instead of plain function calls in C++.

EXPORTED API SYMBOLS

   Symbols.ev      for libev proper
   Symbols.event   for the libevent emulation

This can also be used to rename all public symbols to avoid clashes with multiple versions of libev linked together (which is obviously bad in itself, but sometimes it is inconvenient to avoid this).

A sed command like this will create wrapper "#define"'s that you need to include before including ev.h:

   <Symbols.ev sed -e "s/.*/#define & myprefix_&/" >wrap.h

This would create a file wrap.h which essentially looks like this:

   #define ev_backend     myprefix_ev_backend
   #define ev_check_start myprefix_ev_check_start
   #define ev_check_stop  myprefix_ev_check_stop
   ...

EXAMPLES

The usage in rxvt-unicode is simpler. It has a ev_cpp.h header file that everybody includes and which overrides some configure choices:

   #define EV_FEATURES 8
   #define EV_USE_SELECT 1
   #define EV_PREPARE_ENABLE 1
   #define EV_IDLE_ENABLE 1
   #define EV_SIGNAL_ENABLE 1
   #define EV_CHILD_ENABLE 1
   #define EV_USE_STDEXCEPT 0
   #define EV_CONFIG_H <config.h>

   #include "ev++.h"

And a ev_cpp.C implementation file that contains libev proper and is compiled:

   #include "ev_cpp.h"
   #include "ev.c"

THREADS AND COROUTINES

All libev functions are reentrant and thread-safe unless explicitly documented otherwise, but libev implements no locking itself. This means that you can use as many loops as you want in parallel, as long as there are no concurrent calls into any libev function with the same loop parameter ("ev_default_*" calls have an implicit default loop parameter, of course): libev guarantees that different event loops share no data structures that need any locking.

Or to put it differently: calls with different loop parameters can be done concurrently from multiple threads, calls with the same loop parameter must be done serially (but can be done from different threads, as long as only one thread ever is inside a call at any point in time, e.g. by using a mutex per loop).

Specifically to support threads (and signal handlers), libev implements so-called "ev_async" watchers, which allow some limited form of concurrency on the same event loop, namely waking it up "from the outside".

If you want to know which design (one loop, locking, or multiple loops without or something else still) is best for your problem, then I cannot help you, but here is some generic advice:

• most applications have a main thread: use the default libev loop in that thread, or create a separate thread running only the default loop.

  This helps integrating other libraries or software modules that use libev themselves and don't care/know about threading.

• one loop per thread is usually a good model.

  Doing this is almost never wrong, sometimes a better-performance model exists, but it is always a good start.

• other models exist, such as the leader/follower pattern, where one loop is handed through multiple threads in a kind of round-robin fashion.

  Choosing a model is hard - look around, learn, know that usually you can do better than you currently do :-)

• often you need to talk to some other thread which blocks in the event loop.

  "ev_async" watchers can be used to wake them up from other threads safely (or from signal contexts...).

  An example use would be to communicate signals or other events that only work in the default loop by registering the signal watcher with the default loop and triggering an "ev_async" watcher from the default loop watcher callback into the event loop interested in the signal.

See also "THREAD LOCKING EXAMPLE".

COROUTINES

Libev is very accommodating to coroutines ("cooperative threads"): libev fully supports nesting calls to its functions from different coroutines (e.g. you can call "ev_run" on the same loop from two different coroutines, and switch freely between both coroutines running the loop, as long as you don't confuse yourself). The only exception is that you must not do this from "ev_periodic" reschedule callbacks.

Care has been taken to ensure that libev does not keep local state inside "ev_run", and other calls do not usually allow for coroutine switches as they do not call any callbacks.

COMPILER WARNINGS

However, these are unavoidable for many reasons. For one, each compiler has different warnings, and each user has different tastes regarding warning options. "Warn-free" code therefore cannot be a goal except when targeting a specific compiler and compiler-version.

Another reason is that some compiler warnings require elaborate workarounds, or other changes to the code that make it less clear and less maintainable.

And of course, some compiler warnings are just plain stupid, or simply wrong (because they don't actually warn about the condition their message seems to warn about). For example, certain older gcc versions had some warnings that resulted in an extreme number of false positives. These have been fixed, but some people still insist on making code warn-free with such buggy versions.

While libev is written to generate as few warnings as possible, "warn-free" code is not a goal, and it is recommended not to build libev with any compiler warnings enabled unless you are prepared to cope with them (e.g. by ignoring them). Remember that warnings are just that: warnings, not errors, or proof of bugs.

VALGRIND

If you think you found a bug (memory leak, uninitialised data access etc.) in libev, then check twice: If valgrind reports something like:

   ==2274==    definitely lost: 0 bytes in 0 blocks.
   ==2274==      possibly lost: 0 bytes in 0 blocks.
   ==2274==    still reachable: 256 bytes in 1 blocks.

Then there is no memory leak, just as memory accounted to global variables is not a memleak - the memory is still being referenced, and didn't leak.

Similarly, under some circumstances, valgrind might report kernel bugs as if it were a bug in libev (e.g. in realloc or in the poll backend, although an acceptable workaround has been found here), or it might be confused.

Keep in mind that valgrind is a very good tool, but only a tool. Don't make it into some kind of religion.

If you are unsure about something, feel free to contact the mailing list with the full valgrind report and an explanation on why you think this is a bug in libev (best check the archives, too :). However, don't be annoyed when you get a brisk "this is no bug" answer and take the chance of learning how to interpret valgrind properly.

If you need, for some reason, empty reports from valgrind for your project I suggest using suppression lists.

GNU/LINUX 32 BIT LIMITATIONS

That means that libev compiled in the default environment doesn't support files larger than 2GiB or so, which mainly affects "ev_stat" watchers.

Unfortunately, many programs try to work around this GNU/Linux issue by enabling the large file API, which makes them incompatible with the standard libev compiled for their system.

Likewise, libev cannot enable the large file API itself as this would suddenly make it incompatible to the default compile time environment, i.e. all programs not using special compile switches.

OS/X AND DARWIN BUGS

"kqueue" is buggy

The kqueue syscall is broken in all known versions - most versions support only sockets, many support pipes.

Libev tries to work around this by not using "kqueue" by default on this rotten platform, but of course you can still ask for it when creating a loop - embedding a socket-only kqueue loop into a select-based one is probably going to work well.

"poll" is buggy

Instead of fixing "kqueue", Apple replaced their (working) "poll" implementation by something calling "kqueue" internally around the 10.5.6 release, so now "kqueue" and "poll" are broken.

Libev tries to work around this by not using "poll" by default on this rotten platform, but of course you can still ask for it when creating a loop.

"select" is buggy

All that's left is "select", and of course Apple found a way to fuck this one up as well: On OS/X, "select" actively limits the number of file descriptors you can pass in to 1024 - your program suddenly crashes when you use more.

There is an undocumented "workaround" for this - defining "_DARWIN_UNLIMITED_SELECT", which libev tries to use, so select should work on OS/X.

SOLARIS PROBLEMS AND WORKAROUNDS

The default compile environment on Solaris is unfortunately so thread-unsafe that you can't even use components/libraries compiled without "-D_REENTRANT" in a threaded program, which, of course, isn't defined by default. A valid, if stupid, implementation choice.

If you want to use libev in threaded environments you have to make sure it's compiled with "_REENTRANT" defined.

Event port backend

The scalable event interface for Solaris is called "event ports". Unfortunately, this mechanism is very buggy in all major releases. If you run into high CPU usage, your program freezes or you get a large number of spurious wakeups, make sure you have all the relevant and latest kernel patches applied. No, I don't know which ones, but there are multiple ones to apply, and afterwards, event ports actually work great.

If you can't get it to work, you can try running the program by setting the environment variable "LIBEV_FLAGS=3" to only allow "poll" and "select" backends.

AIX POLL BUG

WIN32 PLATFORM LIMITATIONS AND WORKAROUNDS

Win32 doesn't support any of the standards (e.g. POSIX) that libev requires, and its I/O model is fundamentally incompatible with the POSIX model. Libev still offers limited functionality on this platform in the form of the "EVBACKEND_SELECT" backend, and only supports socket descriptors. This only applies when using Win32 natively, not when using e.g. cygwin. Actually, it only applies to the microsofts own compilers, as every compiler comes with a slightly differently broken/incompatible environment.

Lifting these limitations would basically require the full re-implementation of the I/O system. If you are into this kind of thing, then note that glib does exactly that for you in a very portable way (note also that glib is the slowest event library known to man).

There is no supported compilation method available on windows except embedding it into other applications.

Sensible signal handling is officially unsupported by Microsoft - libev tries its best, but under most conditions, signals will simply not work.

Not a libev limitation but worth mentioning: windows apparently doesn't accept large writes: instead of resulting in a partial write, windows will either accept everything or return "ENOBUFS" if the buffer is too large, so make sure you only write small amounts into your sockets (less than a megabyte seems safe, but this apparently depends on the amount of memory available).

Due to the many, low, and arbitrary limits on the win32 platform and the abysmal performance of winsockets, using a large number of sockets is not recommended (and not reasonable). If your program needs to use more than a hundred or so sockets, then likely it needs to use a totally different implementation for windows, as libev offers the POSIX readiness notification model, which cannot be implemented efficiently on windows (due to Microsoft monopoly games).

A typical way to use libev under windows is to embed it (see the embedding section for details) and use the following evwrap.h header file instead of ev.h:

   #define EV_STANDALONE              /* keeps ev from requiring config.h */
   #define EV_SELECT_IS_WINSOCKET 1   /* configure libev for windows select */

   #include "ev.h"

And compile the following evwrap.c file into your project (make sure you do not compile the ev.c or any other embedded source files!):

   #include "evwrap.h"
   #include "ev.c"

The winsocket "select" function

The winsocket "select" function doesn't follow POSIX in that it requires socket handles and not socket file descriptors (it is also extremely buggy). This makes select very inefficient, and also requires a mapping from file descriptors to socket handles (the Microsoft C runtime provides the function "_open_osfhandle" for this). See the discussion of the "EV_SELECT_USE_FD_SET", "EV_SELECT_IS_WINSOCKET" and "EV_FD_TO_WIN32_HANDLE" preprocessor symbols for more info.

The configuration for a "naked" win32 using the Microsoft runtime libraries and raw winsocket select is:

   #define EV_USE_SELECT 1
   #define EV_SELECT_IS_WINSOCKET 1   /* forces EV_SELECT_USE_FD_SET, too */

Note that winsockets handling of fd sets is O(n), so you can easily get a complexity in the O(n²) range when using win32.

Limited number of file descriptors

Windows has numerous arbitrary (and low) limits on things.

Early versions of winsocket's select only supported waiting for a maximum of 64 handles (probably owning to the fact that all windows kernels can only wait for 64 things at the same time internally; Microsoft recommends spawning a chain of threads and wait for 63 handles and the previous thread in each. Sounds great!).

Newer versions support more handles, but you need to define "FD_SETSIZE" to some high number (e.g. 2048) before compiling the winsocket select call (which might be in libev or elsewhere, for example, perl and many other interpreters do their own select emulation on windows).

Another limit is the number of file descriptors in the Microsoft runtime libraries, which by default is 64 (there must be a hidden 64 fetish or something like this inside Microsoft). You can increase this by calling "_setmaxstdio", which can increase this limit to 2048 (another arbitrary limit), but is broken in many versions of the Microsoft runtime libraries. This might get you to about 512 or 2048 sockets (depending on windows version and/or the phase of the moon). To get more, you need to wrap all I/O functions and provide your own fd management, but the cost of calling select (O(n²)) will likely make this unworkable.

PORTABILITY REQUIREMENTS

"void (*)(ev_watcher_type *, int revents)" must have compatible calling conventions regardless of "ev_watcher_type *".

Libev assumes not only that all watcher pointers have the same internal structure (guaranteed by POSIX but not by ISO C for example), but it also assumes that the same (machine) code can be used to call any watcher callback: The watcher callbacks have different type signatures, but libev calls them using an "ev_watcher *" internally.

null pointers and integer zero are represented by 0 bytes

Libev uses "memset" to initialise structs and arrays to 0 bytes, and relies on this setting pointers and integers to null.

pointer accesses must be thread-atomic

Accessing a pointer value must be atomic, it must both be readable and writable in one piece - this is the case on all current architectures.

"sig_atomic_t volatile" must be thread-atomic as well

The type "sig_atomic_t volatile" (or whatever is defined as "EV_ATOMIC_T") must be atomic with respect to accesses from different threads. This is not part of the specification for "sig_atomic_t", but is believed to be sufficiently portable.

"sigprocmask" must work in a threaded environment

Libev uses "sigprocmask" to temporarily block signals. This is not allowed in a threaded program ("pthread_sigmask" has to be used). Typical pthread implementations will either allow "sigprocmask" in the "main thread" or will block signals process-wide, both behaviours would be compatible with libev. Interaction between "sigprocmask" and "pthread_sigmask" could complicate things, however.

The most portable way to handle signals is to block signals in all threads except the initial one, and run the signal handling loop in the initial thread as well.

"long" must be large enough for common memory allocation sizes

To improve portability and simplify its API, libev uses "long" internally instead of "size_t" when allocating its data structures. On non-POSIX systems (Microsoft...) this might be unexpectedly low, but is still at least 31 bits everywhere, which is enough for hundreds of millions of watchers.

"double" must hold a time value in seconds with enough accuracy

The type "double" is used to represent timestamps. It is required to have at least 51 bits of mantissa (and 9 bits of exponent), which is good enough for at least into the year 4000 with millisecond accuracy (the design goal for libev). This requirement is overfulfilled by implementations using IEEE 754, which is basically all existing ones.

With IEEE 754 doubles, you get microsecond accuracy until at least the year 2255 (and millisecond accuracy till the year 287396 - by then, libev is either obsolete or somebody patched it to use "long double" or something like that, just kidding).

If you know of other additional requirements drop me a note.